#include "cpu-amd64.h"
static gint lmf_tls_offset = -1;
+static gint lmf_addr_tls_offset = -1;
static gint appdomain_tls_offset = -1;
static gint thread_tls_offset = -1;
#ifdef MONO_XEN_OPT
-/* TRUE by default until we add runtime detection of Xen */
static gboolean optimize_for_xen = TRUE;
#else
#define optimize_for_xen 0
static gboolean use_sse2 = !MONO_ARCH_USE_FPSTACK;
-const char * const amd64_desc [OP_LAST];
-static const char*const * ins_spec = amd64_desc;
-
#define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
#define IS_IMM32(val) ((((guint64)val) >> 32) == 0)
+#define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
+
#ifdef PLATFORM_WIN32
/* Under windows, the default pinvoke calling convention is stdcall */
#define CALLCONV_IS_STDCALL(call_conv) (((call_conv) == MONO_CALL_STDCALL) || ((call_conv) == MONO_CALL_DEFAULT))
#define CALLCONV_IS_STDCALL(call_conv) ((call_conv) == MONO_CALL_STDCALL)
#endif
+/* This mutex protects architecture specific caches */
+#define mono_mini_arch_lock() EnterCriticalSection (&mini_arch_mutex)
+#define mono_mini_arch_unlock() LeaveCriticalSection (&mini_arch_mutex)
+static CRITICAL_SECTION mini_arch_mutex;
+
#define ARGS_OFFSET 16
#define GP_SCRATCH_REG AMD64_R11
x86_patch (code, (unsigned char*)target);
}
+void
+mono_amd64_patch (unsigned char* code, gpointer target)
+{
+ amd64_patch (code, target);
+}
+
typedef enum {
ArgInIReg,
ArgInFloatSSEReg,
* Draft Version 0.23" document for more information.
*/
static CallInfo*
-get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
+get_call_info (MonoCompile *cfg, MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
{
guint32 i, gr, fr;
MonoType *ret_type;
guint32 stack_size = 0;
CallInfo *cinfo;
- cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
+ if (mp)
+ cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
+ else
+ cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
gr = 0;
fr = 0;
/* return value */
{
ret_type = mono_type_get_underlying_type (sig->ret);
+ ret_type = mini_get_basic_type_from_generic (cfg, ret_type);
switch (ret_type->type) {
case MONO_TYPE_BOOLEAN:
case MONO_TYPE_I1:
continue;
}
ptype = mono_type_get_underlying_type (sig->params [i]);
+ ptype = mini_get_basic_type_from_generic (cfg, ptype);
switch (ptype->type) {
case MONO_TYPE_BOOLEAN:
case MONO_TYPE_I1:
mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
{
int k;
- CallInfo *cinfo = get_call_info (csig, FALSE);
+ CallInfo *cinfo = get_call_info (NULL, NULL, csig, FALSE);
guint32 args_size = cinfo->stack_usage;
/* The arguments are saved to a stack area in mono_arch_instrument_prolog */
void
mono_arch_cpu_init (void)
{
+#ifndef _MSC_VER
guint16 fpcw;
/* spec compliance requires running with double precision */
fpcw |= X86_FPCW_PREC_DOUBLE;
__asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
__asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
+#else
+ _control87 (_PC_53, MCW_PC);
+#endif
+}
+
+/*
+ * Initialize architecture specific code.
+ */
+void
+mono_arch_init (void)
+{
+ InitializeCriticalSection (&mini_arch_mutex);
+}
+
+/*
+ * Cleanup architecture specific code.
+ */
+void
+mono_arch_cleanup (void)
+{
+ DeleteCriticalSection (&mini_arch_mutex);
}
/*
return use_sse2;
}
-static gboolean
-is_regsize_var (MonoType *t) {
- if (t->byref)
- return TRUE;
- t = mono_type_get_underlying_type (t);
- switch (t->type) {
- case MONO_TYPE_I4:
- case MONO_TYPE_U4:
- case MONO_TYPE_I:
- case MONO_TYPE_U:
- case MONO_TYPE_PTR:
- case MONO_TYPE_FNPTR:
- return TRUE;
- case MONO_TYPE_OBJECT:
- case MONO_TYPE_STRING:
- case MONO_TYPE_CLASS:
- case MONO_TYPE_SZARRAY:
- case MONO_TYPE_ARRAY:
- return TRUE;
- case MONO_TYPE_GENERICINST:
- if (!mono_type_generic_inst_is_valuetype (t))
- return TRUE;
- return FALSE;
- case MONO_TYPE_VALUETYPE:
- return FALSE;
- }
- return FALSE;
-}
-
GList *
mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
{
(ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
continue;
- /* we dont allocate I1 to registers because there is no simply way to sign extend
- * 8bit quantities in caller saved registers on x86 */
- if (is_regsize_var (ins->inst_vtype) || (ins->inst_vtype->type == MONO_TYPE_BOOLEAN) ||
- (ins->inst_vtype->type == MONO_TYPE_U1) || (ins->inst_vtype->type == MONO_TYPE_U2)||
- (ins->inst_vtype->type == MONO_TYPE_I2) || (ins->inst_vtype->type == MONO_TYPE_CHAR)) {
+ if (mono_is_regsize_var (ins->inst_vtype)) {
g_assert (MONO_VARINFO (cfg, i)->reg == -1);
g_assert (i == vmv->idx);
vars = g_list_prepend (vars, vmv);
{
MonoMethodSignature *sig;
MonoMethodHeader *header;
- int i;
+ int i, locals_size;
CallInfo *cinfo;
if (cfg->arch.omit_fp_computed)
sig = mono_method_signature (cfg->method);
- cinfo = get_call_info (sig, FALSE);
+ if (!cfg->arch.cinfo)
+ cfg->arch.cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
+ cinfo = cfg->arch.cinfo;
/*
* FIXME: Remove some of the restrictions.
}
}
- if (cfg->num_varinfo > 10000) {
+ locals_size = 0;
+ for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
+ MonoInst *ins = cfg->varinfo [i];
+ int ialign;
+
+ locals_size += mono_type_size (ins->inst_vtype, &ialign);
+ }
+
+ if ((cfg->num_varinfo > 10000) || (locals_size >= (1 << 15))) {
/* Avoid hitting the stack_alloc_size < (1 << 16) assertion in emit_epilog () */
cfg->arch.omit_fp = FALSE;
}
-
- g_free (cinfo);
}
GList *
sig = mono_method_signature (cfg->method);
- cinfo = get_call_info (sig, FALSE);
+ cinfo = cfg->arch.cinfo;
mono_arch_compute_omit_fp (cfg);
/* Reserve stack space for saving LMF + argument regs */
guint32 size = sizeof (MonoLMF);
- if (lmf_tls_offset == -1)
+ if (lmf_addr_tls_offset == -1)
/* Need to save argument regs too */
size += (AMD64_NREG * 8) + (8 * 8);
//printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
}
}
- g_free (offsets);
offset += locals_stack_size;
if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
}
for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
- inst = cfg->varinfo [i];
+ inst = cfg->args [i];
if (inst->opcode != OP_REGVAR) {
ArgInfo *ainfo = &cinfo->args [i];
gboolean inreg = TRUE;
}
cfg->stack_offset = offset;
-
- g_free (cinfo);
}
void
sig = mono_method_signature (cfg->method);
- cinfo = get_call_info (sig, FALSE);
+ if (!cfg->arch.cinfo)
+ cfg->arch.cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
+ cinfo = cfg->arch.cinfo;
if (cinfo->ret.storage == ArgValuetypeInReg)
cfg->ret_var_is_local = TRUE;
-
- g_free (cinfo);
}
static void
sig = call->signature;
n = sig->param_count + sig->hasthis;
- cinfo = get_call_info (sig, sig->pinvoke);
+ cinfo = get_call_info (cfg, cfg->mempool, sig, sig->pinvoke);
for (i = 0; i < n; ++i) {
ainfo = cinfo->args + i;
cfg->param_area = MAX (cfg->param_area, call->stack_usage);
cfg->flags |= MONO_CFG_HAS_CALLS;
- g_free (cinfo);
-
return call;
}
amd64_call_reg (code, GP_SCRATCH_REG);
}
}
-
- return code;
-}
-
-static inline guint8*
-emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
-{
- mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
-
- return emit_call_body (cfg, code, patch_type, data);
+
+ return code;
+}
+
+static inline guint8*
+emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
+{
+ mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
+
+ return emit_call_body (cfg, code, patch_type, data);
+}
+
+static inline int
+store_membase_imm_to_store_membase_reg (int opcode)
+{
+ switch (opcode) {
+ case OP_STORE_MEMBASE_IMM:
+ return OP_STORE_MEMBASE_REG;
+ case OP_STOREI4_MEMBASE_IMM:
+ return OP_STOREI4_MEMBASE_REG;
+ case OP_STOREI8_MEMBASE_IMM:
+ return OP_STOREI8_MEMBASE_REG;
+ }
+
+ return -1;
+}
+
+#define INST_IGNORES_CFLAGS(opcode) (!(((opcode) == OP_ADC) || ((opcode) == OP_ADC_IMM) || ((opcode) == OP_IADC) || ((opcode) == OP_IADC_IMM) || ((opcode) == OP_SBB) || ((opcode) == OP_SBB_IMM) || ((opcode) == OP_ISBB) || ((opcode) == OP_ISBB_IMM)))
+
+/*
+ * peephole_pass_1:
+ *
+ * Perform peephole opts which should/can be performed before local regalloc
+ */
+static void
+peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
+{
+ MonoInst *ins, *last_ins = NULL;
+ ins = bb->code;
+
+ while (ins) {
+
+ switch (ins->opcode) {
+ case OP_ADD_IMM:
+ case OP_IADD_IMM:
+ case OP_LADD_IMM:
+ if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS) && (ins->inst_imm > 0)) {
+ /*
+ * X86_LEA is like ADD, but doesn't have the
+ * sreg1==dreg restriction. inst_imm > 0 is needed since LEA sign-extends
+ * its operand to 64 bit.
+ */
+ ins->opcode = OP_X86_LEA_MEMBASE;
+ ins->inst_basereg = ins->sreg1;
+ /* Fall through */
+ }
+ else
+ break;
+ case CEE_XOR:
+ if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
+ MonoInst *ins2;
+
+ /*
+ * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
+ * the latter has length 2-3 instead of 6 (reverse constant
+ * propagation). These instruction sequences are very common
+ * in the initlocals bblock.
+ */
+ for (ins2 = ins->next; ins2; ins2 = ins2->next) {
+ if (((ins2->opcode == OP_STORE_MEMBASE_IMM) || (ins2->opcode == OP_STOREI4_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_IMM) || (ins2->opcode == OP_STORE_MEMBASE_IMM)) && (ins2->inst_imm == 0)) {
+ ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
+ ins2->sreg1 = ins->dreg;
+ } else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_REG) || (ins2->opcode == OP_STORE_MEMBASE_REG)) {
+ /* Continue */
+ } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
+ NULLIFY_INS (ins2);
+ /* Continue */
+ } else {
+ break;
+ }
+ }
+ }
+ break;
+ case OP_COMPARE_IMM:
+ /* OP_COMPARE_IMM (reg, 0)
+ * -->
+ * OP_AMD64_TEST_NULL (reg)
+ */
+ if (!ins->inst_imm)
+ ins->opcode = OP_AMD64_TEST_NULL;
+ break;
+ case OP_ICOMPARE_IMM:
+ if (!ins->inst_imm)
+ ins->opcode = OP_X86_TEST_NULL;
+ break;
+ case OP_AMD64_ICOMPARE_MEMBASE_IMM:
+ /*
+ * OP_STORE_MEMBASE_REG reg, offset(basereg)
+ * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
+ * -->
+ * OP_STORE_MEMBASE_REG reg, offset(basereg)
+ * OP_COMPARE_IMM reg, imm
+ *
+ * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
+ */
+ if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ ins->opcode = OP_ICOMPARE_IMM;
+ ins->sreg1 = last_ins->sreg1;
+
+ /* check if we can remove cmp reg,0 with test null */
+ if (!ins->inst_imm)
+ ins->opcode = OP_X86_TEST_NULL;
+ }
+
+ break;
+ case OP_LOAD_MEMBASE:
+ case OP_LOADI4_MEMBASE:
+ /*
+ * Note: if reg1 = reg2 the load op is removed
+ *
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_MOVE reg1, reg2
+ */
+ if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG
+ || last_ins->opcode == OP_STORE_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ if (ins->dreg == last_ins->sreg1) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
+ } else {
+ //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
+ ins->opcode = OP_MOVE;
+ ins->sreg1 = last_ins->sreg1;
+ }
+
+ /*
+ * Note: reg1 must be different from the basereg in the second load
+ * Note: if reg1 = reg2 is equal then second load is removed
+ *
+ * OP_LOAD_MEMBASE offset(basereg), reg1
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_LOAD_MEMBASE offset(basereg), reg1
+ * OP_MOVE reg1, reg2
+ */
+ } if (last_ins && (last_ins->opcode == OP_LOADI4_MEMBASE
+ || last_ins->opcode == OP_LOAD_MEMBASE) &&
+ ins->inst_basereg != last_ins->dreg &&
+ ins->inst_basereg == last_ins->inst_basereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+
+ if (ins->dreg == last_ins->dreg) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
+ } else {
+ ins->opcode = OP_MOVE;
+ ins->sreg1 = last_ins->dreg;
+ }
+
+ //g_assert_not_reached ();
+
+#if 0
+ /*
+ * OP_STORE_MEMBASE_IMM imm, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg
+ * -->
+ * OP_STORE_MEMBASE_IMM imm, offset(basereg)
+ * OP_ICONST reg, imm
+ */
+ } else if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_IMM
+ || last_ins->opcode == OP_STORE_MEMBASE_IMM) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
+ ins->opcode = OP_ICONST;
+ ins->inst_c0 = last_ins->inst_imm;
+ g_assert_not_reached (); // check this rule
+#endif
+ }
+ break;
+ case OP_LOADI1_MEMBASE:
+ /*
+ * Note: if reg1 = reg2 the load op is removed
+ *
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_MOVE reg1, reg2
+ */
+ if (last_ins && (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ if (ins->dreg == last_ins->sreg1) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
+ } else {
+ //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
+ ins->opcode = OP_MOVE;
+ ins->sreg1 = last_ins->sreg1;
+ }
+ }
+ break;
+ case OP_LOADI2_MEMBASE:
+ /*
+ * Note: if reg1 = reg2 the load op is removed
+ *
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_MOVE reg1, reg2
+ */
+ if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ if (ins->dreg == last_ins->sreg1) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
+ } else {
+ //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
+ ins->opcode = OP_MOVE;
+ ins->sreg1 = last_ins->sreg1;
+ }
+ }
+ break;
+ case CEE_CONV_I4:
+ case CEE_CONV_U4:
+ case OP_MOVE:
+ case OP_FMOVE:
+ /*
+ * Removes:
+ *
+ * OP_MOVE reg, reg
+ */
+ if (ins->dreg == ins->sreg1) {
+ if (last_ins)
+ last_ins->next = ins->next;
+ else
+ bb->code = ins->next;
+ ins = ins->next;
+ continue;
+ }
+ /*
+ * Removes:
+ *
+ * OP_MOVE sreg, dreg
+ * OP_MOVE dreg, sreg
+ */
+ if (last_ins && last_ins->opcode == OP_MOVE &&
+ ins->sreg1 == last_ins->dreg &&
+ ins->dreg == last_ins->sreg1) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
+ }
+ break;
+ }
+ last_ins = ins;
+ ins = ins->next;
+ }
+ bb->last_ins = last_ins;
}
-/* FIXME: Add more instructions */
-#define INST_IGNORES_CFLAGS(ins) (((ins)->opcode == CEE_BR) || ((ins)->opcode == OP_STORE_MEMBASE_IMM) || ((ins)->opcode == OP_STOREI8_MEMBASE_REG) || ((ins)->opcode == OP_MOVE) || ((ins)->opcode == OP_ICONST) || ((ins)->opcode == OP_I8CONST) || ((ins)->opcode == OP_LOAD_MEMBASE))
-
static void
peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
{
case OP_I8CONST:
/* reg = 0 -> XOR (reg, reg) */
/* XOR sets cflags on x86, so we cant do it always */
- if (ins->inst_c0 == 0 && (ins->next && INST_IGNORES_CFLAGS (ins->next))) {
- ins->opcode = CEE_XOR;
+ if (ins->inst_c0 == 0 && (!ins->next || (ins->next && INST_IGNORES_CFLAGS (ins->next->opcode)))) {
+ ins->opcode = OP_LXOR;
ins->sreg1 = ins->dreg;
ins->sreg2 = ins->dreg;
+ /* Fall through */
+ }
+ else
+ break;
+ case CEE_XOR:
+ case OP_LXOR:
+ if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
+ MonoInst *ins2;
+
+ /*
+ * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
+ * the latter has length 2-3 instead of 6 (reverse constant
+ * propagation). These instruction sequences are very common
+ * in the initlocals bblock.
+ */
+ for (ins2 = ins->next; ins2; ins2 = ins2->next) {
+ if (((ins2->opcode == OP_STORE_MEMBASE_IMM) || (ins2->opcode == OP_STOREI4_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_IMM) || (ins2->opcode == OP_STORE_MEMBASE_IMM)) && (ins2->inst_imm == 0)) {
+ ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
+ ins2->sreg1 = ins->dreg;
+ } else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_REG) || (ins2->opcode == OP_STORE_MEMBASE_REG)) {
+ /* Continue */
+ } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
+ NULLIFY_INS (ins2);
+ /* Continue */
+ } else {
+ break;
+ }
+ }
}
break;
+ case OP_IADD_IMM:
+ if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_INC_REG;
+ break;
+ case OP_ISUB_IMM:
+ if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_DEC_REG;
+ break;
case OP_MUL_IMM:
/* remove unnecessary multiplication with 1 */
if (ins->inst_imm == 1) {
#endif
}
break;
- case OP_LOADU1_MEMBASE:
case OP_LOADI1_MEMBASE:
/*
* Note: if reg1 = reg2 the load op is removed
}
}
break;
- case OP_LOADU2_MEMBASE:
case OP_LOADI2_MEMBASE:
/*
* Note: if reg1 = reg2 the load op is removed
case CEE_CONV_I4:
case CEE_CONV_U4:
case OP_MOVE:
+ case OP_FMOVE:
/*
* Removes:
*
*/
if (ins->dreg == ins->sreg1) {
if (last_ins)
- last_ins->next = ins->next;
+ last_ins->next = ins->next;
+ else
+ bb->code = ins->next;
ins = ins->next;
continue;
}
MonoInst *ins, *temp, *last_ins = NULL;
ins = bb->code;
- if (bb->max_ireg > cfg->rs->next_vireg)
- cfg->rs->next_vireg = bb->max_ireg;
- if (bb->max_freg > cfg->rs->next_vfreg)
- cfg->rs->next_vfreg = bb->max_freg;
+ if (bb->max_vreg > cfg->rs->next_vreg)
+ cfg->rs->next_vreg = bb->max_vreg;
/*
* FIXME: Need to add more instructions, but the current machine
}
bb->last_ins = last_ins;
- bb->max_ireg = cfg->rs->next_vireg;
- bb->max_freg = cfg->rs->next_vfreg;
+ bb->max_vreg = cfg->rs->next_vreg;
}
static const int
X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
};
-static int
-opcode_to_x86_cond (int opcode)
-{
- switch (opcode) {
- case OP_IBEQ:
- return X86_CC_EQ;
- case OP_IBNE_UN:
- return X86_CC_NE;
- case OP_IBLT:
- return X86_CC_LT;
- case OP_IBLT_UN:
- return X86_CC_LT;
- case OP_IBGT:
- return X86_CC_GT;
- case OP_IBGT_UN:
- return X86_CC_GT;
- case OP_IBGE:
- return X86_CC_GE;
- case OP_IBGE_UN:
- return X86_CC_GE;
- case OP_IBLE:
- return X86_CC_LE;
- case OP_IBLE_UN:
- return X86_CC_LE;
- case OP_COND_EXC_IOV:
- return X86_CC_O;
- case OP_COND_EXC_IC:
- return X86_CC_C;
- default:
- g_assert_not_reached ();
- }
+/* Maps CMP_... constants to X86_CC_... constants */
+static const int
+cc_table [] = {
+ X86_CC_EQ, X86_CC_NE, X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT,
+ X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT
+};
- return -1;
-}
+static const int
+cc_signed_table [] = {
+ TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
+ FALSE, FALSE, FALSE, FALSE
+};
/*#include "cprop.c"*/
mono_arch_lowering_pass (cfg, bb);
+ if (cfg->opt & MONO_OPT_PEEPHOLE)
+ peephole_pass_1 (cfg, bb);
+
mono_local_regalloc (cfg, bb);
}
case OP_VCALL:
case OP_VCALL_REG:
case OP_VCALL_MEMBASE:
- cinfo = get_call_info (((MonoCallInst*)ins)->signature, FALSE);
+ cinfo = get_call_info (cfg, cfg->mempool, ((MonoCallInst*)ins)->signature, FALSE);
if (cinfo->ret.storage == ArgValuetypeInReg) {
/* Pop the destination address from the stack */
amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
}
}
}
- g_free (cinfo);
break;
}
sig = mono_method_signature (method);
- cinfo = get_call_info (sig, FALSE);
+ cinfo = cfg->arch.cinfo;
/* This is the opposite of the code in emit_prolog */
for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
ArgInfo *ainfo = cinfo->args + i;
MonoType *arg_type;
- inst = cfg->varinfo [i];
+ inst = cfg->args [i];
if (sig->hasthis && (i == 0))
arg_type = &mono_defaults.object_class->byval_arg;
}
}
- g_free (cinfo);
-
return code;
}
while (ins) {
offset = code - cfg->native_code;
- max_len = ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
+ max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
if (offset > (cfg->code_size - max_len - 16)) {
cfg->code_size *= 2;
case OP_AMD64_ICOMPARE_REG_MEMBASE:
amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
break;
- case CEE_BREAK:
+ case OP_BREAK:
amd64_breakpoint (code);
break;
case OP_ADDCC:
}
case CEE_DIV:
case OP_LDIV:
- amd64_cdq (code);
- amd64_div_reg (code, ins->sreg2, TRUE);
- break;
- case CEE_DIV_UN:
- case OP_LDIV_UN:
- amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
- amd64_div_reg (code, ins->sreg2, FALSE);
- break;
case CEE_REM:
case OP_LREM:
- amd64_cdq (code);
- amd64_div_reg (code, ins->sreg2, TRUE);
+ /* Regalloc magic makes the div/rem cases the same */
+ if (ins->sreg2 == AMD64_RDX) {
+ amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
+ amd64_cdq (code);
+ amd64_div_membase (code, AMD64_RSP, -8, TRUE);
+ } else {
+ amd64_cdq (code);
+ amd64_div_reg (code, ins->sreg2, TRUE);
+ }
break;
+ case CEE_DIV_UN:
+ case OP_LDIV_UN:
case CEE_REM_UN:
case OP_LREM_UN:
- amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
- amd64_div_reg (code, ins->sreg2, FALSE);
+ if (ins->sreg2 == AMD64_RDX) {
+ amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
+ amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
+ amd64_div_membase (code, AMD64_RSP, -8, FALSE);
+ } else {
+ amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
+ amd64_div_reg (code, ins->sreg2, FALSE);
+ }
+ break;
+ case OP_IDIV:
+ case OP_IREM:
+ if (ins->sreg2 == AMD64_RDX) {
+ amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
+ amd64_cdq_size (code, 4);
+ amd64_div_membase_size (code, AMD64_RSP, -8, TRUE, 4);
+ } else {
+ amd64_cdq_size (code, 4);
+ amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
+ }
+ break;
+ case OP_IDIV_UN:
+ case OP_IREM_UN:
+ if (ins->sreg2 == AMD64_RDX) {
+ amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
+ amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
+ amd64_div_membase_size (code, AMD64_RSP, -8, FALSE, 4);
+ } else {
+ amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
+ amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
+ }
break;
case OP_LMUL_OVF:
amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
amd64_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
break;
case CEE_XOR:
+ case OP_LXOR:
amd64_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
break;
case OP_XOR_IMM:
EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
break;
}
- case OP_IDIV:
- amd64_cdq_size (code, 4);
- amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
- break;
- case OP_IDIV_UN:
- amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
- amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
- break;
- case OP_IREM:
- amd64_cdq_size (code, 4);
- amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
- break;
- case OP_IREM_UN:
- amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
- amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
- break;
case OP_ICOMPARE:
amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
break;
case OP_IBGT:
case OP_IBGE:
case OP_IBLE:
- EMIT_COND_BRANCH (ins, opcode_to_x86_cond (ins->opcode), TRUE);
- break;
case OP_IBNE_UN:
case OP_IBLT_UN:
case OP_IBGT_UN:
case OP_IBGE_UN:
case OP_IBLE_UN:
- EMIT_COND_BRANCH (ins, opcode_to_x86_cond (ins->opcode), FALSE);
- break;
- case OP_COND_EXC_IOV:
- EMIT_COND_SYSTEM_EXCEPTION (opcode_to_x86_cond (ins->opcode),
- TRUE, ins->inst_p1);
- break;
- case OP_COND_EXC_IC:
- EMIT_COND_SYSTEM_EXCEPTION (opcode_to_x86_cond (ins->opcode),
- FALSE, ins->inst_p1);
+ EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
break;
case CEE_NOT:
amd64_not_reg (code, ins->sreg1);
}
break;
}
- case CEE_JMP: {
+ case OP_JMP: {
/*
* Note: this 'frame destruction' logic is useful for tail calls, too.
* Keep in sync with the code in emit_epilog.
case CEE_RET:
amd64_ret (code);
break;
- case CEE_THROW: {
+ case OP_THROW: {
amd64_mov_reg_reg (code, AMD64_RDI, ins->sreg1, 8);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
(gpointer)"mono_arch_throw_exception");
/* Restore stack alignment */
amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
break;
+
case OP_LABEL:
ins->inst_c0 = code - cfg->native_code;
break;
- case CEE_BR:
+ case OP_NOP:
+ break;
+ case OP_BR:
//g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
//if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
//break;
break;
case OP_CEQ:
case OP_ICEQ:
- amd64_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
- amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CLT:
case OP_ICLT:
- amd64_set_reg (code, X86_CC_LT, ins->dreg, TRUE);
- amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
- case OP_CLT_UN:
- case OP_ICLT_UN:
- amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
- amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CGT:
case OP_ICGT:
- amd64_set_reg (code, X86_CC_GT, ins->dreg, TRUE);
- amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
+ case OP_CLT_UN:
+ case OP_ICLT_UN:
case OP_CGT_UN:
case OP_ICGT_UN:
- amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
+ amd64_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
break;
case OP_COND_EXC_EQ:
case OP_COND_EXC_GE_UN:
case OP_COND_EXC_LE:
case OP_COND_EXC_LE_UN:
+ EMIT_COND_SYSTEM_EXCEPTION (cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], ins->inst_p1);
+ break;
case OP_COND_EXC_OV:
case OP_COND_EXC_NO:
case OP_COND_EXC_C:
EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
(ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
break;
+ case OP_COND_EXC_IOV:
+ case OP_COND_EXC_IC:
+ EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_IEQ],
+ (ins->opcode < OP_COND_EXC_INE_UN), ins->inst_p1);
+ break;
case CEE_BEQ:
case CEE_BNE_UN:
case CEE_BLT:
case CEE_BGE_UN:
case CEE_BLE:
case CEE_BLE_UN:
- EMIT_COND_BRANCH (ins, branch_cc_table [ins->opcode - CEE_BEQ], (ins->opcode < CEE_BNE_UN));
+ EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
break;
/* floating point opcodes */
br = code; x86_branch8 (code, X86_CC_GEZ, 0, TRUE);
/* add correction constant mn */
- x86_fld80_mem (code, mn);
+ x86_fld80_mem (code, (gssize)mn);
x86_fld80_membase (code, AMD64_RSP, 0);
amd64_fp_op_reg (code, X86_FADD, 1, TRUE);
x86_fst80_membase (code, AMD64_RSP, 0);
case OP_FBGT:
case OP_FBGT_UN:
if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
- EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
+ if (ins->opcode == OP_FBGT) {
+ guchar *br1;
+
+ /* skip branch if C1=1 */
+ br1 = code;
+ x86_branch8 (code, X86_CC_P, 0, FALSE);
+ /* branch if (C0 | C3) = 1 */
+ EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
+ amd64_patch (br1, code);
+ break;
+ } else {
+ EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
+ }
break;
}
amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
break;
- case CEE_CKFINITE: {
+ case OP_CKFINITE: {
if (use_sse2) {
/* Transfer value to the fp stack */
amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 16);
* hack to overcome limits in x86 reg allocator
* (req: dreg == eax and sreg2 != eax and breg != eax)
*/
+ /* The pushes invalidate rsp */
+ if ((breg == AMD64_RAX) || (breg == AMD64_RSP)) {
+ amd64_mov_reg_reg (code, AMD64_R11, breg, 8);
+ breg = AMD64_R11;
+ }
+
if (ins->dreg != AMD64_RAX)
amd64_push_reg (code, AMD64_RAX);
sreg2 = AMD64_RDX;
}
- if (breg == AMD64_RAX) {
- amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
- breg = AMD64_R11;
- }
-
amd64_mov_reg_membase (code, AMD64_RAX, breg, ins->inst_offset, size);
br [0] = code; amd64_prefix (code, X86_LOCK_PREFIX);
}
}
+/*
+ * This macro is used for testing whenever the unwinder works correctly at every point
+ * where an async exception can happen.
+ */
+/* This will generate a SIGSEGV at the given point in the code */
+#define async_exc_point(code) do { \
+ if (mono_inject_async_exc_method && mono_method_desc_full_match (mono_inject_async_exc_method, cfg->method)) { \
+ if (cfg->arch.async_point_count == mono_inject_async_exc_pos) \
+ amd64_mov_reg_mem (code, AMD64_RAX, 0, 4); \
+ cfg->arch.async_point_count ++; \
+ } \
+} while (0)
+
guint8 *
mono_arch_emit_prolog (MonoCompile *cfg)
{
CallInfo *cinfo;
gint32 lmf_offset = cfg->arch.lmf_offset;
- cfg->code_size = MAX (((MonoMethodNormal *)method)->header->code_size * 4, 512);
+ cfg->code_size = MAX (((MonoMethodNormal *)method)->header->code_size * 4, 1024);
+
+ if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
+ cfg->code_size += 512;
+
code = cfg->native_code = g_malloc (cfg->code_size);
/* Amount of stack space allocated by register saving code */
* - save callee saved regs using moves
*/
+ async_exc_point (code);
+
if (!cfg->arch.omit_fp) {
amd64_push_reg (code, AMD64_RBP);
+ async_exc_point (code);
amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof (gpointer));
+ async_exc_point (code);
}
/* Save callee saved registers */
if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
amd64_push_reg (code, i);
pos += sizeof (gpointer);
+ async_exc_point (code);
}
}
guint32 remaining_size = alloc_size;
while (remaining_size >= 0x1000) {
amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
+ async_exc_point (code);
amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
remaining_size -= 0x1000;
}
- if (remaining_size)
+ if (remaining_size) {
amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, remaining_size);
+ async_exc_point (code);
+ }
#else
amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, alloc_size);
+ async_exc_point (code);
#endif
}
amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp), AMD64_RBP, 8);
/* Save sp */
amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
- /* Save method */
- /* FIXME: add a relocation for this */
- if (IS_IMM32 (cfg->method))
- amd64_mov_membase_imm (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), (guint64)cfg->method, 8);
- else {
- amd64_mov_reg_imm (code, AMD64_R11, cfg->method);
- amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, 8);
- }
+ /* Skip method (only needed for trampoline LMF frames) */
/* Save callee saved regs */
amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, 8);
amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, 8);
if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
amd64_mov_membase_reg (code, AMD64_RSP, save_area_offset, i, 8);
save_area_offset += 8;
+ async_exc_point (code);
}
}
if (ins->opcode == OP_LABEL)
ins->inst_c1 = max_offset;
- max_offset += ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
+ max_offset += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
ins = ins->next;
}
}
sig = mono_method_signature (method);
pos = 0;
- cinfo = get_call_info (sig, FALSE);
+ cinfo = cfg->arch.cinfo;
if (sig->ret->type != MONO_TYPE_VOID) {
if ((cinfo->ret.storage == ArgInIReg) && (cfg->ret->opcode != OP_REGVAR)) {
ArgInfo *ainfo = cinfo->args + i;
gint32 stack_offset;
MonoType *arg_type;
- inst = cfg->varinfo [i];
+ inst = cfg->args [i];
if (sig->hasthis && (i == 0))
arg_type = &mono_defaults.object_class->byval_arg;
* The call might clobber argument registers, but they are already
* saved to the stack/global regs.
*/
- if (lmf_tls_offset != -1) {
+ if (lmf_addr_tls_offset != -1) {
guint8 *buf;
- code = emit_tls_get ( code, AMD64_RAX, lmf_tls_offset);
+ code = emit_tls_get ( code, AMD64_RAX, lmf_addr_tls_offset);
amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
buf = code;
x86_branch8 (code, X86_CC_NE, 0, 0);
}
if (method->save_lmf) {
- if (lmf_tls_offset != -1) {
- /* Load lmf quicky using the FS register */
- code = emit_tls_get (code, AMD64_RAX, lmf_tls_offset);
- }
- else {
- /*
- * The call might clobber argument registers, but they are already
- * saved to the stack/global regs.
+ if ((lmf_tls_offset != -1) && !optimize_for_xen) {
+ /*
+ * Optimized version which uses the mono_lmf TLS variable instead of indirection
+ * through the mono_lmf_addr TLS variable.
*/
+ /* %rax = previous_lmf */
+ x86_prefix (code, X86_FS_PREFIX);
+ amd64_mov_reg_mem (code, AMD64_RAX, lmf_tls_offset, 8);
+
+ /* Save previous_lmf */
+ amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_RAX, 8);
+ /* Set new lmf */
+ amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
+ x86_prefix (code, X86_FS_PREFIX);
+ amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
+ } else {
+ if (lmf_addr_tls_offset != -1) {
+ /* Load lmf quicky using the FS register */
+ code = emit_tls_get (code, AMD64_RAX, lmf_addr_tls_offset);
+ }
+ else {
+ /*
+ * The call might clobber argument registers, but they are already
+ * saved to the stack/global regs.
+ */
+ code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
+ (gpointer)"mono_get_lmf_addr");
+ }
- code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
- (gpointer)"mono_get_lmf_addr");
+ /* Save lmf_addr */
+ amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, 8);
+ /* Save previous_lmf */
+ amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, 8);
+ amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, 8);
+ /* Set new lmf */
+ amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
+ amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, 8);
}
-
- /* Save lmf_addr */
- amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, 8);
- /* Save previous_lmf */
- amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, 8);
- amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, 8);
- /* Set new lmf */
- amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
- amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, 8);
}
-
- g_free (cinfo);
-
if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
- /* the code restoring the registers must be kept in sync with CEE_JMP */
+ /* the code restoring the registers must be kept in sync with OP_JMP */
pos = 0;
if (method->save_lmf) {
- /* Restore previous lmf */
- amd64_mov_reg_membase (code, AMD64_RCX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
- amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 8);
- amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, 8);
+ if ((lmf_tls_offset != -1) && !optimize_for_xen) {
+ /*
+ * Optimized version which uses the mono_lmf TLS variable instead of indirection
+ * through the mono_lmf_addr TLS variable.
+ */
+ /* reg = previous_lmf */
+ amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
+ x86_prefix (code, X86_FS_PREFIX);
+ amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
+ } else {
+ /* Restore previous lmf */
+ amd64_mov_reg_membase (code, AMD64_RCX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
+ amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 8);
+ amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, 8);
+ }
/* Restore caller saved regs */
if (cfg->used_int_regs & (1 << AMD64_RBP)) {
}
/* Load returned vtypes into registers if needed */
- cinfo = get_call_info (mono_method_signature (method), FALSE);
+ cinfo = cfg->arch.cinfo;
if (cinfo->ret.storage == ArgValuetypeInReg) {
ArgInfo *ainfo = &cinfo->ret;
MonoInst *inst = cfg->ret;
}
}
}
- g_free (cinfo);
if (cfg->arch.omit_fp) {
if (cfg->arch.stack_alloc_size)
} else {
amd64_leave (code);
}
+ async_exc_point (code);
amd64_ret (code);
cfg->code_len = code - cfg->native_code;
pos = cfg->native_code + patch_info->ip.i;
- if (use_sse2)
- *(guint32*)(pos + 4) = (guint8*)code - pos - 8;
- else
+
+ if (use_sse2) {
+ if (IS_REX (pos [1]))
+ *(guint32*)(pos + 5) = (guint8*)code - pos - 9;
+ else
+ *(guint32*)(pos + 4) = (guint8*)code - pos - 8;
+ } else {
*(guint32*)(pos + 3) = (guint8*)code - pos - 7;
+ }
if (patch_info->type == MONO_PATCH_INFO_R8) {
*(double*)code = *(double*)patch_info->data.target;
/* Allocate a new area on the stack and save arguments there */
sig = mono_method_signature (cfg->method);
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
n = sig->param_count + sig->hasthis;
amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, stack_area);
for (i = 0; i < n; ++i) {
- inst = cfg->varinfo [i];
+ inst = cfg->args [i];
if (inst->opcode == OP_REGVAR)
amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), inst->dreg, 8);
amd64_mov_reg_reg (code, AMD64_RSI, AMD64_RSP, 8);
code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
- if (enable_arguments) {
+ if (enable_arguments)
amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, stack_area);
- g_free (cinfo);
- }
-
return code;
}
return amd64_is_imm32 (imm);
}
-#define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
-
/*
* Determine whenever the trap whose info is in SIGINFO is caused by
* integer overflow.
mono_arch_get_vcall_slot_addr (guint8* code, gpointer *regs)
{
guint32 reg;
- guint32 disp;
+ gint32 disp;
guint8 rex = 0;
/* go to the start of the call instruction
* really careful about the ordering of the cases. Longer sequences
* come first.
*/
- if ((code [-1] == 0x8b) && (amd64_modrm_mod (code [0]) == 0x2) && (code [5] == 0xff) && (amd64_modrm_reg (code [6]) == 0x2) && (amd64_modrm_mod (code [6]) == 0x0)) {
+#ifdef MONO_ARCH_HAVE_IMT
+ if ((code [-2] == 0x41) && (code [-1] == 0xbb) && (code [4] == 0xff) && (x86_modrm_mod (code [5]) == 1) && (x86_modrm_reg (code [5]) == 2) && ((signed char)code [6] < 0)) {
+ /* IMT-based interface calls: with MONO_ARCH_IMT_REG == r11
+ * 41 bb 14 f8 28 08 mov $0x828f814,%r11d
+ * ff 50 fc call *0xfffffffc(%rax)
+ */
+ reg = amd64_modrm_rm (code [5]);
+ disp = (signed char)code [6];
+ /* R10 is clobbered by the IMT thunk code */
+ g_assert (reg != AMD64_R10);
+ }
+#else
+ if (0) {
+ }
+#endif
+ else if ((code [-1] == 0x8b) && (amd64_modrm_mod (code [0]) == 0x2) && (code [5] == 0xff) && (amd64_modrm_reg (code [6]) == 0x2) && (amd64_modrm_mod (code [6]) == 0x0)) {
/*
* This is a interface call
* 48 8b 80 f0 e8 ff ff mov 0xffffffffffffe8f0(%rax),%rax
rex = code [4];
reg = amd64_modrm_rm (code [6]);
disp = 0;
- }
- else if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
+ /* R10 is clobbered by the IMT thunk code */
+ g_assert (reg != AMD64_R10);
+ } else if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
/* call OFFSET(%rip) */
disp = *(guint32*)(code + 3);
return (gpointer*)(code + disp + 7);
if (IS_REX (code [0]))
rex = code [0];
reg = amd64_modrm_rm (code [2]);
- disp = *(guint32*)(code + 3);
- //printf ("B: [%%r%d+0x%x]\n", reg, disp);
+ disp = *(gint32*)(code + 3);
+ /* R10 is clobbered by the IMT thunk code */
+ g_assert (reg != AMD64_R10);
}
else if (code [2] == 0xe8) {
/* call <ADDR> */
if (IS_REX (code [3]))
rex = code [3];
reg = amd64_modrm_rm (code [5]);
- disp = *(guint8*)(code + 6);
+ disp = *(gint8*)(code + 6);
//printf ("B: [%%r%d+0x%x]\n", reg, disp);
}
else if ((code [5] == 0xff) && (amd64_modrm_reg (code [6]) == 0x2) && (amd64_modrm_mod (code [6]) == 0x0)) {
return (gpointer)(((guint64)(regs [reg])) + disp);
}
-gpointer*
-mono_arch_get_delegate_method_ptr_addr (guint8* code, gpointer *regs)
+gpointer
+mono_arch_get_this_arg_from_call (MonoMethodSignature *sig, gssize *regs, guint8 *code)
{
- guint32 reg;
- guint32 disp;
+ if (MONO_TYPE_ISSTRUCT (sig->ret))
+ return (gpointer)regs [AMD64_RSI];
+ else
+ return (gpointer)regs [AMD64_RDI];
+}
- code -= 10;
+#define MAX_ARCH_DELEGATE_PARAMS 10
+
+gpointer
+mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
+{
+ guint8 *code, *start;
+ int i;
+
+ if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
+ return NULL;
+
+ /* FIXME: Support more cases */
+ if (MONO_TYPE_ISSTRUCT (sig->ret))
+ return NULL;
+
+ if (has_target) {
+ static guint8* cached = NULL;
+ mono_mini_arch_lock ();
+ if (cached) {
+ mono_mini_arch_unlock ();
+ return cached;
+ }
+
+ start = code = mono_global_codeman_reserve (64);
+
+ /* Replace the this argument with the target */
+ amd64_mov_reg_reg (code, AMD64_RAX, AMD64_RDI, 8);
+ amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, target), 8);
+ amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
- if (IS_REX (code [0]) && (code [1] == 0x8b) && (code [3] == 0x48) && (code [4] == 0x8b) && (code [5] == 0x40) && (code [7] == 0x48) && (code [8] == 0xff) && (code [9] == 0xd0)) {
- /* mov REG, %rax; mov <OFFSET>(%rax), %rax; call *%rax */
- reg = amd64_rex_b (code [0]) + amd64_modrm_rm (code [2]);
- disp = code [6];
+ g_assert ((code - start) < 64);
- if (reg == AMD64_RAX)
+ cached = start;
+ mono_mini_arch_unlock ();
+ } else {
+ static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
+ for (i = 0; i < sig->param_count; ++i)
+ if (!mono_is_regsize_var (sig->params [i]))
+ return NULL;
+ if (sig->param_count > 4)
return NULL;
- else
- return (gpointer*)(((guint64)(regs [reg])) + disp);
+
+ mono_mini_arch_lock ();
+ code = cache [sig->param_count];
+ if (code) {
+ mono_mini_arch_unlock ();
+ return code;
+ }
+
+ start = code = mono_global_codeman_reserve (64);
+
+ if (sig->param_count == 0) {
+ amd64_jump_membase (code, AMD64_RDI, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
+ } else {
+ /* We have to shift the arguments left */
+ amd64_mov_reg_reg (code, AMD64_RAX, AMD64_RDI, 8);
+ for (i = 0; i < sig->param_count; ++i)
+ amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
+
+ amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
+ }
+ g_assert ((code - start) < 64);
+
+ cache [sig->param_count] = start;
+
+ mono_mini_arch_unlock ();
}
- return NULL;
+ return start;
}
/*
optimize_for_xen = access ("/proc/xen", F_OK) == 0;
#endif
appdomain_tls_offset = mono_domain_get_tls_offset ();
- lmf_tls_offset = mono_get_lmf_tls_offset ();
+ lmf_tls_offset = mono_get_lmf_tls_offset ();
+ lmf_addr_tls_offset = mono_get_lmf_addr_tls_offset ();
thread_tls_offset = mono_thread_get_tls_offset ();
}
}
mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
{
MonoCallInst *call = (MonoCallInst*)inst;
- CallInfo * cinfo = get_call_info (inst->signature, FALSE);
+ CallInfo * cinfo = get_call_info (cfg, cfg->mempool, inst->signature, FALSE);
if (vt_reg != -1) {
MonoInst *vtarg;
mono_call_inst_add_outarg_reg (cfg, call, this->dreg, cinfo->args [0].reg, FALSE);
}
+}
- g_free (cinfo);
+#ifdef MONO_ARCH_HAVE_IMT
+
+#define CMP_SIZE (6 + 1)
+#define CMP_REG_REG_SIZE (4 + 1)
+#define BR_SMALL_SIZE 2
+#define BR_LARGE_SIZE 6
+#define MOV_REG_IMM_SIZE 10
+#define MOV_REG_IMM_32BIT_SIZE 6
+#define JUMP_REG_SIZE (2 + 1)
+
+static int
+imt_branch_distance (MonoIMTCheckItem **imt_entries, int start, int target)
+{
+ int i, distance = 0;
+ for (i = start; i < target; ++i)
+ distance += imt_entries [i]->chunk_size;
+ return distance;
+}
+
+/*
+ * LOCKING: called with the domain lock held
+ */
+gpointer
+mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count)
+{
+ int i;
+ int size = 0;
+ guint8 *code, *start;
+ gboolean vtable_is_32bit = ((long)(vtable) == (long)(int)(long)(vtable));
+
+ for (i = 0; i < count; ++i) {
+ MonoIMTCheckItem *item = imt_entries [i];
+ if (item->is_equals) {
+ if (item->check_target_idx) {
+ if (!item->compare_done) {
+ if (amd64_is_imm32 (item->method))
+ item->chunk_size += CMP_SIZE;
+ else
+ item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
+ }
+ if (vtable_is_32bit)
+ item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
+ else
+ item->chunk_size += MOV_REG_IMM_SIZE;
+ item->chunk_size += BR_SMALL_SIZE + JUMP_REG_SIZE;
+ } else {
+ if (vtable_is_32bit)
+ item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
+ else
+ item->chunk_size += MOV_REG_IMM_SIZE;
+ item->chunk_size += JUMP_REG_SIZE;
+ /* with assert below:
+ * item->chunk_size += CMP_SIZE + BR_SMALL_SIZE + 1;
+ */
+ }
+ } else {
+ if (amd64_is_imm32 (item->method))
+ item->chunk_size += CMP_SIZE;
+ else
+ item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
+ item->chunk_size += BR_LARGE_SIZE;
+ imt_entries [item->check_target_idx]->compare_done = TRUE;
+ }
+ size += item->chunk_size;
+ }
+ code = mono_code_manager_reserve (domain->code_mp, size);
+ start = code;
+ for (i = 0; i < count; ++i) {
+ MonoIMTCheckItem *item = imt_entries [i];
+ item->code_target = code;
+ if (item->is_equals) {
+ if (item->check_target_idx) {
+ if (!item->compare_done) {
+ if (amd64_is_imm32 (item->method))
+ amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->method);
+ else {
+ amd64_mov_reg_imm (code, AMD64_R10, item->method);
+ amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, AMD64_R10);
+ }
+ }
+ item->jmp_code = code;
+ amd64_branch8 (code, X86_CC_NE, 0, FALSE);
+ amd64_mov_reg_imm (code, AMD64_R11, & (vtable->vtable [item->vtable_slot]));
+ amd64_jump_membase (code, AMD64_R11, 0);
+ } else {
+ /* enable the commented code to assert on wrong method */
+#if 0
+ if (amd64_is_imm32 (item->method))
+ amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->method);
+ else {
+ amd64_mov_reg_imm (code, AMD64_R10, item->method);
+ amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, AMD64_R10);
+ }
+ item->jmp_code = code;
+ amd64_branch8 (code, X86_CC_NE, 0, FALSE);
+ amd64_mov_reg_imm (code, AMD64_R11, & (vtable->vtable [item->vtable_slot]));
+ amd64_jump_membase (code, AMD64_R11, 0);
+ amd64_patch (item->jmp_code, code);
+ amd64_breakpoint (code);
+ item->jmp_code = NULL;
+#else
+ amd64_mov_reg_imm (code, AMD64_R11, & (vtable->vtable [item->vtable_slot]));
+ amd64_jump_membase (code, AMD64_R11, 0);
+#endif
+ }
+ } else {
+ if (amd64_is_imm32 (item->method))
+ amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->method);
+ else {
+ amd64_mov_reg_imm (code, AMD64_R10, item->method);
+ amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, AMD64_R10);
+ }
+ item->jmp_code = code;
+ if (x86_is_imm8 (imt_branch_distance (imt_entries, i, item->check_target_idx)))
+ x86_branch8 (code, X86_CC_GE, 0, FALSE);
+ else
+ x86_branch32 (code, X86_CC_GE, 0, FALSE);
+ }
+ g_assert (code - item->code_target <= item->chunk_size);
+ }
+ /* patch the branches to get to the target items */
+ for (i = 0; i < count; ++i) {
+ MonoIMTCheckItem *item = imt_entries [i];
+ if (item->jmp_code) {
+ if (item->check_target_idx) {
+ amd64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
+ }
+ }
+ }
+
+ mono_stats.imt_thunks_size += code - start;
+ g_assert (code - start <= size);
+
+ return start;
+}
+
+MonoMethod*
+mono_arch_find_imt_method (gpointer *regs, guint8 *code)
+{
+ /*
+ * R11 is clobbered by the trampoline code, so we have to retrieve the method
+ * from the code.
+ * 41 bb c0 f7 89 00 mov $0x89f7c0,%r11d
+ * ff 90 68 ff ff ff callq *0xffffffffffffff68(%rax)
+ */
+ /* Similar to get_vcall_slot_addr () */
+
+ /* Find the start of the call instruction */
+ code -= 7;
+ if ((code [-2] == 0x41) && (code [-1] == 0xbb) && (code [4] == 0xff) && (x86_modrm_mod (code [5]) == 1) && (x86_modrm_reg (code [5]) == 2) && ((signed char)code [6] < 0)) {
+ /* IMT-based interface calls
+ * 41 bb 14 f8 28 08 mov $0x828f814,%r11
+ * ff 50 fc call *0xfffffffc(%rax)
+ */
+ code += 4;
+ } else if ((code [1] == 0xff) && (amd64_modrm_reg (code [2]) == 0x2) && (amd64_modrm_mod (code [2]) == 0x2)) {
+ /* call *[reg+disp32] */
+ code += 1;
+ } else if ((code [4] == 0xff) && (amd64_modrm_reg (code [5]) == 0x2) && (amd64_modrm_mod (code [5]) == 0x1)) {
+ /* call *[reg+disp8] */
+ code += 4;
+ } else
+ g_assert_not_reached ();
+
+ /* Find the start of the mov instruction */
+ code -= 10;
+ if (code [0] == 0x49 && code [1] == 0xbb) {
+ return (MonoMethod*)*(gssize*)(code + 2);
+ } else if (code [3] == 0x4d && code [4] == 0x8b && code [5] == 0x1d) {
+ /* mov <OFFSET>(%rip),%r11 */
+ return (MonoMethod*)*(gssize*)(code + 10 + *(guint32*)(code + 6));
+ } else if (code [4] == 0x41 && code [5] == 0xbb) {
+ return (MonoMethod*)(gssize)*(guint32*)(code + 6);
+ } else {
+ int i;
+
+ printf ("Unknown call sequence: ");
+ for (i = -10; i < 20; ++i)
+ printf ("%x ", code [i]);
+ g_assert_not_reached ();
+ return NULL;
+ }
}
+MonoObject*
+mono_arch_find_this_argument (gpointer *regs, MonoMethod *method)
+{
+ return mono_arch_get_this_arg_from_call (mono_method_signature (method), (gssize*)regs, NULL);
+}
+#endif
+
MonoInst*
mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
{