#include "mini.h"
#include <string.h>
#include <math.h>
+#ifdef HAVE_UNISTD_H
#include <unistd.h>
+#endif
#include <mono/metadata/appdomain.h>
#include <mono/metadata/debug-helpers.h>
#include <mono/metadata/threads.h>
#include <mono/metadata/profiler-private.h>
+#include <mono/metadata/mono-debug.h>
#include <mono/utils/mono-math.h>
#include "trace.h"
static gboolean is_win32 = FALSE;
#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 ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
#define ARGS_OFFSET 8
#define NOT_IMPLEMENTED g_assert_not_reached ()
+MonoBreakpointInfo
+mono_breakpoint_info [MONO_BREAKPOINT_ARRAY_SIZE];
+
const char*
-mono_arch_regname (int reg) {
+mono_arch_regname (int reg)
+{
switch (reg) {
case X86_EAX: return "%eax";
case X86_EBX: return "%ebx";
case X86_ECX: return "%ecx";
case X86_EDX: return "%edx";
- case X86_ESP: return "%esp"; case X86_EBP: return "%ebp";
+ case X86_ESP: return "%esp";
+ case X86_EBP: return "%ebp";
case X86_EDI: return "%edi";
case X86_ESI: return "%esi";
}
}
const char*
-mono_arch_fregname (int reg) {
- return "unknown";
+mono_arch_fregname (int reg)
+{
+ switch (reg) {
+ case 0:
+ return "%fr0";
+ case 1:
+ return "%fr1";
+ case 2:
+ return "%fr2";
+ case 3:
+ return "%fr3";
+ case 4:
+ return "%fr4";
+ case 5:
+ return "%fr5";
+ case 6:
+ return "%fr6";
+ case 7:
+ return "%fr7";
+ default:
+ return "unknown";
+ }
}
typedef enum {
static void
-add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
+add_valuetype (MonoGenericSharingContext *gsctx, MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
gboolean is_return,
guint32 *gr, guint32 *fr, guint32 *stack_size)
{
if (sig->pinvoke)
size = mono_type_native_stack_size (&klass->byval_arg, NULL);
else
- size = mono_type_stack_size (&klass->byval_arg, NULL);
+ size = mini_type_stack_size (gsctx, &klass->byval_arg, NULL);
#ifdef SMALL_STRUCTS_IN_REGS
if (sig->pinvoke && is_return) {
* For x86 win32, see ???.
*/
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;
int n = sig->hasthis + sig->param_count;
guint32 stack_size = 0;
CallInfo *cinfo;
+ MonoGenericSharingContext *gsctx = cfg ? cfg->generic_sharing_context : NULL;
- 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 (gsctx, ret_type);
switch (ret_type->type) {
case MONO_TYPE_BOOLEAN:
case MONO_TYPE_I1:
case MONO_TYPE_VALUETYPE: {
guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
- add_valuetype (sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
+ add_valuetype (gsctx, sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
if (cinfo->ret.storage == ArgOnStack)
/* The caller passes the address where the value is stored */
add_general (&gr, &stack_size, &cinfo->ret);
continue;
}
ptype = mono_type_get_underlying_type (sig->params [i]);
+ ptype = mini_get_basic_type_from_generic (gsctx, ptype);
switch (ptype->type) {
case MONO_TYPE_BOOLEAN:
case MONO_TYPE_I1:
}
/* Fall through */
case MONO_TYPE_VALUETYPE:
- add_valuetype (sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
+ add_valuetype (gsctx, sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
break;
case MONO_TYPE_TYPEDBYREF:
stack_size += sizeof (MonoTypedRef);
int offset = 8;
CallInfo *cinfo;
- cinfo = get_call_info (csig, FALSE);
+ cinfo = get_call_info (NULL, NULL, csig, FALSE);
if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
frame_size += sizeof (gpointer);
size = mono_type_native_stack_size (csig->params [k], &align);
else {
int ialign;
- size = mono_type_stack_size (csig->params [k], &ialign);
+ size = mini_type_stack_size (NULL, csig->params [k], &ialign);
align = ialign;
}
#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);
+}
+
/*
* This function returns the optimizations supported on this cpu.
*/
*exclude_mask |= MONO_OPT_FCMOV;
} else
*exclude_mask |= MONO_OPT_CMOV;
+ if (edx & (1 << 26))
+ opts |= MONO_OPT_SSE2;
+ else
+ *exclude_mask |= MONO_OPT_SSE2;
}
return opts;
}
header = mono_method_get_header (cfg->method);
sig = mono_method_signature (cfg->method);
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
cfg->frame_reg = MONO_ARCH_BASEREG;
offset = 0;
offset &= ~(MONO_ARCH_FRAME_ALIGNMENT - 1);
cfg->stack_offset = offset;
-
- g_free (cinfo);
}
void
sig = mono_method_signature (cfg->method);
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
if (cinfo->ret.storage == ArgValuetypeInReg)
cfg->ret_var_is_local = TRUE;
-
- g_free (cinfo);
}
/* Fixme: we need an alignment solution for enter_method and mono_arch_call_opcode,
call->out_args = arg;
}
+/*
+ * It is expensive to adjust esp for each individual fp argument pushed on the stack
+ * so we try to do it just once when we have multiple fp arguments in a row.
+ * We don't use this mechanism generally because for int arguments the generated code
+ * is slightly bigger and new generation cpus optimize away the dependency chains
+ * created by push instructions on the esp value.
+ * fp_arg_setup is the first argument in the execution sequence where the esp register
+ * is modified.
+ */
+static int
+collect_fp_stack_space (MonoMethodSignature *sig, int start_arg, int *fp_arg_setup)
+{
+ int fp_space = 0;
+ MonoType *t;
+
+ for (; start_arg < sig->param_count; ++start_arg) {
+ t = mono_type_get_underlying_type (sig->params [start_arg]);
+ if (!t->byref && t->type == MONO_TYPE_R8) {
+ fp_space += sizeof (double);
+ *fp_arg_setup = start_arg;
+ } else {
+ break;
+ }
+ }
+ return fp_space;
+}
+
/*
* take the arguments and generate the arch-specific
* instructions to properly call the function in call.
int i, n;
CallInfo *cinfo;
int sentinelpos = 0;
+ int fp_args_space = 0, fp_args_offset = 0, fp_arg_setup = -1;
sig = call->signature;
n = sig->param_count + sig->hasthis;
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
sentinelpos = sig->sentinelpos + (is_virtual ? 1 : 0);
size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
else {
int ialign;
- size = mono_type_stack_size (&in->klass->byval_arg, &ialign);
+ size = mini_type_stack_size (cfg->generic_sharing_context, &in->klass->byval_arg, &ialign);
align = ialign;
}
arg->opcode = OP_OUTARG_VT;
case ArgOnStack:
arg->opcode = OP_OUTARG;
if (!t->byref) {
- if (t->type == MONO_TYPE_R4)
+ if (t->type == MONO_TYPE_R4) {
arg->opcode = OP_OUTARG_R4;
- else
- if (t->type == MONO_TYPE_R8)
- arg->opcode = OP_OUTARG_R8;
+ } else if (t->type == MONO_TYPE_R8) {
+ arg->opcode = OP_OUTARG_R8;
+ /* we store in the upper bits of backen.arg_info the needed
+ * esp adjustment and in the lower bits the offset from esp
+ * where the arg needs to be stored
+ */
+ if (!fp_args_space) {
+ fp_args_space = collect_fp_stack_space (sig, i - sig->hasthis, &fp_arg_setup);
+ fp_args_offset = fp_args_space;
+ }
+ arg->backend.arg_info = fp_args_space - fp_args_offset;
+ fp_args_offset -= sizeof (double);
+ if (i - sig->hasthis == fp_arg_setup) {
+ arg->backend.arg_info |= fp_args_space << 16;
+ }
+ if (fp_args_offset == 0) {
+ /* the allocated esp stack is finished:
+ * prepare for an eventual second run of fp args
+ */
+ fp_args_space = 0;
+ }
+ }
}
break;
default:
}
#endif
- g_free (cinfo);
-
return call;
}
return code;
}
-/* FIXME: Add more instructions */
-#define INST_IGNORES_CFLAGS(ins) (((ins)->opcode == OP_BR) || ((ins)->opcode == OP_STORE_MEMBASE_IMM) || ((ins)->opcode == OP_STOREI4_MEMBASE_REG))
+#define INST_IGNORES_CFLAGS(opcode) (!(((opcode) == OP_ADC) || ((opcode) == OP_IADC) || ((opcode) == OP_ADC_IMM) || ((opcode) == OP_IADC_IMM) || ((opcode) == OP_SBB) || ((opcode) == OP_ISBB) || ((opcode) == OP_SBB_IMM) || ((opcode) == OP_ISBB_IMM)))
+/*
+ * peephole_pass_1:
+ *
+ * Perform peephole opts which should/can be performed before local regalloc
+ */
static void
-peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
+peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
{
MonoInst *ins, *last_ins = NULL;
ins = bb->code;
while (ins) {
-
switch (ins->opcode) {
- case OP_ICONST:
- /* 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;
- ins->sreg1 = ins->dreg;
- ins->sreg2 = ins->dreg;
- }
+ case OP_IADD_IMM:
+ case OP_ADD_IMM:
+ if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS)) {
+ /*
+ * X86_LEA is like ADD, but doesn't have the
+ * sreg1==dreg restriction.
+ */
+ ins->opcode = OP_X86_LEA_MEMBASE;
+ ins->inst_basereg = ins->sreg1;
+ } else if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_INC_REG;
break;
- case OP_MUL_IMM:
- /* remove unnecessary multiplication with 1 */
- if (ins->inst_imm == 1) {
- if (ins->dreg != ins->sreg1) {
- ins->opcode = OP_MOVE;
- } else {
- last_ins->next = ins->next;
- ins = ins->next;
- continue;
- }
- }
+ case OP_SUB_IMM:
+ case OP_ISUB_IMM:
+ if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS)) {
+ ins->opcode = OP_X86_LEA_MEMBASE;
+ ins->inst_basereg = ins->sreg1;
+ ins->inst_imm = -ins->inst_imm;
+ } else if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_DEC_REG;
break;
case OP_COMPARE_IMM:
+ case OP_ICOMPARE_IMM:
/* OP_COMPARE_IMM (reg, 0)
* -->
* OP_X86_TEST_NULL (reg)
break;
case CEE_CONV_I4:
case CEE_CONV_U4:
+ case OP_ICONV_TO_I4:
case OP_MOVE:
/*
* Removes:
bb->last_ins = last_ins;
}
+static void
+peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
+{
+ MonoInst *ins, *last_ins = NULL;
+ ins = bb->code;
+
+ while (ins) {
+
+ switch (ins->opcode) {
+ case OP_ICONST:
+ /* reg = 0 -> XOR (reg, reg) */
+ /* XOR sets cflags on x86, so we cant do it always */
+ if (ins->inst_c0 == 0 && (!ins->next || (ins->next && INST_IGNORES_CFLAGS (ins->next->opcode)))) {
+ MonoInst *ins2;
+
+ ins->opcode = OP_IXOR;
+ ins->sreg1 = ins->dreg;
+ ins->sreg2 = ins->dreg;
+
+ /*
+ * Convert succeeding STORE_MEMBASE_IMM 0 ins to STORE_MEMBASE_REG
+ * since it takes 3 bytes instead of 7.
+ */
+ for (ins2 = ins->next; ins2; ins2 = ins2->next) {
+ if ((ins2->opcode == OP_STORE_MEMBASE_IMM) && (ins2->inst_imm == 0)) {
+ ins2->opcode = OP_STORE_MEMBASE_REG;
+ ins2->sreg1 = ins->dreg;
+ }
+ else if ((ins2->opcode == OP_STOREI4_MEMBASE_IMM) && (ins2->inst_imm == 0)) {
+ ins2->opcode = OP_STOREI4_MEMBASE_REG;
+ ins2->sreg1 = ins->dreg;
+ }
+ else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM)) {
+ /* Continue iteration */
+ }
+ else
+ break;
+ }
+ }
+ break;
+ case OP_IADD_IMM:
+ case OP_ADD_IMM:
+ if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_INC_REG;
+ break;
+ case OP_ISUB_IMM:
+ case OP_SUB_IMM:
+ if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_DEC_REG;
+ break;
+ case OP_X86_COMPARE_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_COMPARE_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_LOADU1_MEMBASE:
+ case OP_LOADI1_MEMBASE:
+ /*
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * CONV_I2/U2 reg1, reg2
+ */
+ if (last_ins && X86_IS_BYTE_REG (last_ins->sreg1) &&
+ (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ ins->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? CEE_CONV_I1 : CEE_CONV_U1;
+ ins->sreg1 = last_ins->sreg1;
+ }
+ break;
+ case OP_LOADU2_MEMBASE:
+ case OP_LOADI2_MEMBASE:
+ /*
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * CONV_I2/U2 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) {
+ ins->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? CEE_CONV_I2 : CEE_CONV_U2;
+ ins->sreg1 = last_ins->sreg1;
+ }
+ break;
+ case CEE_CONV_I4:
+ case CEE_CONV_U4:
+ case OP_ICONV_TO_I4:
+ case OP_MOVE:
+ /*
+ * Removes:
+ *
+ * OP_MOVE reg, reg
+ */
+ if (ins->dreg == ins->sreg1) {
+ if (last_ins)
+ last_ins->next = 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;
+ case OP_X86_PUSH_MEMBASE:
+ 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) {
+ ins->opcode = OP_X86_PUSH;
+ ins->sreg1 = last_ins->sreg1;
+ }
+ break;
+ }
+ last_ins = ins;
+ ins = ins->next;
+ }
+ bb->last_ins = last_ins;
+}
+
static const int
branch_cc_table [] = {
X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
};
-/*#include "cprop.c"*/
+/* 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
+};
+
+static const int
+cc_signed_table [] = {
+ TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
+ FALSE, FALSE, FALSE, FALSE
+};
+
void
mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
{
+ if (cfg->opt & MONO_OPT_PEEPHOLE)
+ peephole_pass_1 (cfg, bb);
+
mono_local_regalloc (cfg, bb);
}
static unsigned char*
emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int size, gboolean is_signed)
{
+#define XMM_TEMP_REG 0
+ if (cfg->opt & MONO_OPT_SSE2 && size < 8) {
+ /* optimize by assigning a local var for this use so we avoid
+ * the stack manipulations */
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
+ x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
+ x86_movsd_reg_membase (code, XMM_TEMP_REG, X86_ESP, 0);
+ x86_cvttsd2si (code, dreg, XMM_TEMP_REG);
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
+ if (size == 1)
+ x86_widen_reg (code, dreg, dreg, is_signed, FALSE);
+ else if (size == 2)
+ x86_widen_reg (code, dreg, dreg, is_signed, TRUE);
+ return code;
+ }
x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
x86_fnstcw_membase(code, X86_ESP, 0);
x86_mov_reg_membase (code, dreg, X86_ESP, 0, 2);
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 */
x86_pop_reg (code, X86_ECX);
}
}
}
- g_free (cinfo);
default:
break;
}
sig = mono_method_signature (method);
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
/* This is the opposite of the code in emit_prolog */
}
}
- g_free (cinfo);
-
return code;
}
max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
- if (offset > (cfg->code_size - max_len - 16)) {
+ if (G_UNLIKELY (offset > (cfg->code_size - max_len - 16))) {
cfg->code_size *= 2;
cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
code = cfg->native_code + offset;
mono_jit_stats.code_reallocs++;
}
- mono_debug_record_line_number (cfg, ins, offset);
+ if (cfg->debug_info)
+ mono_debug_record_line_number (cfg, ins, offset);
switch (ins->opcode) {
case OP_BIGMUL:
x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
break;
case CEE_CONV_I1:
+ case OP_SEXT_I1:
x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
break;
case CEE_CONV_I2:
+ case OP_SEXT_I2:
x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
break;
case CEE_CONV_U1:
x86_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
break;
case CEE_XOR:
+ case OP_IXOR:
x86_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
break;
case OP_XOR_IMM:
case CEE_NEG:
x86_neg_reg (code, ins->sreg1);
break;
- case OP_SEXT_I1:
- x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
- break;
- case OP_SEXT_I2:
- x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
- break;
+
case CEE_MUL:
x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
break;
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
#endif
break;
+ case OP_START_HANDLER: {
+ MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
+ x86_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, X86_ESP, 4);
+ break;
+ }
+ case OP_ENDFINALLY: {
+ MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
+ x86_mov_reg_membase (code, X86_ESP, spvar->inst_basereg, spvar->inst_offset, 4);
+ x86_ret (code);
+ break;
+ }
+ case OP_ENDFILTER: {
+ MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
+ x86_mov_reg_membase (code, X86_ESP, spvar->inst_basereg, spvar->inst_offset, 4);
+ /* The local allocator will put the result into EAX */
+ x86_ret (code);
+ break;
+ }
+
case OP_LABEL:
ins->inst_c0 = code - cfg->native_code;
break;
x86_jump_reg (code, ins->sreg1);
break;
case OP_CEQ:
- x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CLT:
- x86_set_reg (code, X86_CC_LT, ins->dreg, TRUE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CLT_UN:
- x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CGT:
- x86_set_reg (code, X86_CC_GT, ins->dreg, TRUE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CGT_UN:
- x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CNE:
- x86_set_reg (code, X86_CC_NE, ins->dreg, TRUE);
+ x86_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
x86_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:
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 */
int alloc_size, pos, max_offset, i;
guint8 *code;
- cfg->code_size = MAX (mono_method_get_header (method)->code_size * 4, 256);
+ cfg->code_size = MAX (mono_method_get_header (method)->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);
x86_push_reg (code, X86_EBP);
pos = 0;
if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
- /* Might need to attach the thread to the JIT */
- if (lmf_tls_offset != -1) {
- guint8 *buf;
+ /* Might need to attach the thread to the JIT or change the domain for the callback */
+ if (appdomain_tls_offset != -1 && lmf_tls_offset != -1) {
+ guint8 *buf, *no_domain_branch;
+ code = emit_tls_get (code, X86_EAX, appdomain_tls_offset);
+ x86_alu_reg_imm (code, X86_CMP, X86_EAX, GPOINTER_TO_UINT (cfg->domain));
+ no_domain_branch = code;
+ x86_branch8 (code, X86_CC_NE, 0, 0);
code = emit_tls_get ( code, X86_EAX, lmf_tls_offset);
x86_test_reg_reg (code, X86_EAX, X86_EAX);
buf = code;
x86_branch8 (code, X86_CC_NE, 0, 0);
+ x86_patch (no_domain_branch, code);
x86_push_imm (code, cfg->domain);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
/* %eax = previous_lmf */
x86_prefix (code, X86_GS_PREFIX);
x86_mov_reg_mem (code, X86_EAX, lmf_tls_offset, 4);
- /* skip method_info + lmf */
- x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
+ /* skip esp + method_info + lmf */
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 12);
/* push previous_lmf */
x86_push_reg (code, X86_EAX);
/* new lmf = ESP */
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
}
- /* Skip method info */
- x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
+ /* Skip esp + method info */
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
/* push lmf */
x86_push_reg (code, X86_EAX);
if (cfg->method->save_lmf)
max_epilog_size += 128;
-
- if (mono_jit_trace_calls != NULL)
- max_epilog_size += 50;
while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
cfg->code_size *= 2;
}
/* Load returned vtypes into registers if needed */
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
if (cinfo->ret.storage == ArgValuetypeInReg) {
for (quad = 0; quad < 2; quad ++) {
switch (cinfo->ret.pair_storage [quad]) {
else
x86_ret (code);
- g_free (cinfo);
-
cfg->code_len = code - cfg->native_code;
g_assert (cfg->code_len < cfg->code_size);
exc_throw_start [nthrows] = code;
}
- x86_push_imm (code, exc_class->type_token);
+ x86_push_imm (code, exc_class->type_token - MONO_TOKEN_TYPE_DEF);
patch_info->data.name = "mono_arch_throw_corlib_exception";
patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
patch_info->ip.i = code - cfg->native_code;
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);
/* add the this argument */
if (this_reg != -1) {
mono_bblock_add_inst (cfg->cbb, vtarg);
}
}
+}
+
+#ifdef MONO_ARCH_HAVE_IMT
+
+// Linear handler, the bsearch head compare is shorter
+//[2 + 4] x86_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
+//[1 + 1] x86_branch8(inst,cond,imm,is_signed)
+// x86_patch(ins,target)
+//[1 + 5] x86_jump_mem(inst,mem)
+
+#define CMP_SIZE 6
+#define BR_SMALL_SIZE 2
+#define BR_LARGE_SIZE 5
+#define JUMP_IMM_SIZE 6
+#define ENABLE_WRONG_METHOD_CHECK 0
+
+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;
+
+ for (i = 0; i < count; ++i) {
+ MonoIMTCheckItem *item = imt_entries [i];
+ if (item->is_equals) {
+ if (item->check_target_idx) {
+ if (!item->compare_done)
+ item->chunk_size += CMP_SIZE;
+ item->chunk_size += BR_SMALL_SIZE + JUMP_IMM_SIZE;
+ } else {
+ item->chunk_size += JUMP_IMM_SIZE;
+#if ENABLE_WRONG_METHOD_CHECK
+ item->chunk_size += CMP_SIZE + BR_SMALL_SIZE + 1;
+#endif
+ }
+ } else {
+ item->chunk_size += CMP_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)
+ x86_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)item->method);
+ item->jmp_code = code;
+ x86_branch8 (code, X86_CC_NE, 0, FALSE);
+ x86_jump_mem (code, & (vtable->vtable [item->vtable_slot]));
+ } else {
+ /* enable the commented code to assert on wrong method */
+#if ENABLE_WRONG_METHOD_CHECK
+ x86_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)item->method);
+ item->jmp_code = code;
+ x86_branch8 (code, X86_CC_NE, 0, FALSE);
+#endif
+ x86_jump_mem (code, & (vtable->vtable [item->vtable_slot]));
+#if ENABLE_WRONG_METHOD_CHECK
+ x86_patch (item->jmp_code, code);
+ x86_breakpoint (code);
+ item->jmp_code = NULL;
+#endif
+ }
+ } else {
+ x86_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)item->method);
+ 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);
+ }
+ }
+ /* 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) {
+ x86_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)
+{
+ return (MonoMethod*) regs [MONO_ARCH_IMT_REG];
+}
+
+MonoObject*
+mono_arch_find_this_argument (gpointer *regs, MonoMethod *method)
+{
+ MonoMethodSignature *sig = mono_method_signature (method);
+ CallInfo *cinfo = get_call_info (NULL, NULL, sig, FALSE);
+ int this_argument_offset;
+ MonoObject *this_argument;
+
+ /*
+ * this is the offset of the this arg from esp as saved at the start of
+ * mono_arch_create_trampoline_code () in tramp-x86.c.
+ */
+ this_argument_offset = 5;
+ if (MONO_TYPE_ISSTRUCT (sig->ret) && (cinfo->ret.storage == ArgOnStack))
+ this_argument_offset++;
+
+ this_argument = * (MonoObject**) (((guint8*) regs [X86_ESP]) + this_argument_offset * sizeof (gpointer));
g_free (cinfo);
+ return this_argument;
}
+#endif
MonoInst*
mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
ins->inst_i1 = args [1];
}
#endif
- } else if (cmethod->klass == mono_defaults.thread_class &&
- strcmp (cmethod->name, "MemoryBarrier") == 0) {
- MONO_INST_NEW (cfg, ins, OP_MEMORY_BARRIER);
- } else if(cmethod->klass->image == mono_defaults.corlib &&
- (strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
- (strcmp (cmethod->klass->name, "Interlocked") == 0)) {
-
- if (strcmp (cmethod->name, "Increment") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
- MonoInst *ins_iconst;
-
- MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
- MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
- ins_iconst->inst_c0 = 1;
-
- ins->inst_i0 = args [0];
- ins->inst_i1 = ins_iconst;
- } else if (strcmp (cmethod->name, "Decrement") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
- MonoInst *ins_iconst;
-
- MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
- MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
- ins_iconst->inst_c0 = -1;
-
- ins->inst_i0 = args [0];
- ins->inst_i1 = ins_iconst;
- } else if (strcmp (cmethod->name, "Exchange") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
- MONO_INST_NEW (cfg, ins, OP_ATOMIC_EXCHANGE_I4);
-
- ins->inst_i0 = args [0];
- ins->inst_i1 = args [1];
- } else if (strcmp (cmethod->name, "Add") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
- MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
-
- ins->inst_i0 = args [0];
- ins->inst_i1 = args [1];
- }
}
return ins;
}
}
-gpointer*
-mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
+gboolean
+mono_breakpoint_clean_code (guint8 *code, guint8 *buf, int size)
{
+ int i;
+ gboolean can_write = TRUE;
+ memcpy (buf, code, size);
+ for (i = 0; i < MONO_BREAKPOINT_ARRAY_SIZE; ++i) {
+ int idx = mono_breakpoint_info_index [i];
+ guint8 *ptr;
+ if (idx < 1)
+ continue;
+ ptr = mono_breakpoint_info [idx].address;
+ if (ptr >= code && ptr < code + size) {
+ guint8 saved_byte = mono_breakpoint_info [idx].saved_byte;
+ can_write = FALSE;
+ /*g_print ("patching %p with 0x%02x (was: 0x%02x)\n", ptr, saved_byte, buf [ptr - code]);*/
+ buf [ptr - code] = saved_byte;
+ }
+ }
+ return can_write;
+}
+
+gpointer
+mono_arch_get_vcall_slot (guint8 *code, gpointer *regs, int *displacement)
+{
+ guint8 buf [8];
guint8 reg = 0;
gint32 disp = 0;
+ mono_breakpoint_clean_code (code - 8, buf, sizeof (buf));
+ code = buf + 8;
+
+ *displacement = 0;
+
/* go to the start of the call instruction
*
* address_byte = (m << 6) | (o << 3) | reg
*/
reg = x86_modrm_rm (code [5]);
disp = 0;
+#ifdef MONO_ARCH_HAVE_IMT
+ } else if ((code [-2] == 0xba) && (code [3] == 0xff) && (x86_modrm_mod (code [4]) == 1) && (x86_modrm_reg (code [4]) == 2) && ((signed char)code [5] < 0)) {
+ /* IMT-based interface calls: with MONO_ARCH_IMT_REG == edx
+ * ba 14 f8 28 08 mov $0x828f814,%edx
+ * ff 50 fc call *0xfffffffc(%eax)
+ */
+ reg = code [4] & 0x07;
+ disp = (signed char)code [5];
+#endif
} else if ((code [1] != 0xe8) && (code [3] == 0xff) && ((code [4] & 0x18) == 0x10) && ((code [4] >> 6) == 1)) {
reg = code [4] & 0x07;
disp = (signed char)code [5];
return NULL;
}
- return (gpointer*)(((gint32)(regs [reg])) + disp);
+ *displacement = disp;
+ return regs [reg];
}
-gpointer*
-mono_arch_get_delegate_method_ptr_addr (guint8* code, gpointer *regs)
+gpointer*
+mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
{
- guint8 reg = 0;
- gint32 disp = 0;
+ gpointer vt;
+ int displacement;
+ vt = mono_arch_get_vcall_slot (code, regs, &displacement);
+ if (!vt)
+ return NULL;
+ return (gpointer*)((char*)vt + displacement);
+}
- code -= 7;
- if ((code [0] == 0x8b) && (x86_modrm_mod (code [1]) == 3) && (x86_modrm_reg (code [1]) == X86_EAX) && (code [2] == 0x8b) && (code [3] == 0x40) && (code [5] == 0xff) && (code [6] == 0xd0)) {
- reg = x86_modrm_rm (code [1]);
- disp = code [4];
+gpointer
+mono_arch_get_this_arg_from_call (MonoMethodSignature *sig, gssize *regs, guint8 *code)
+{
+ guint32 esp = regs [X86_ESP];
+ CallInfo *cinfo;
+ gpointer res;
- if (reg == X86_EAX)
- return NULL;
- else
- return (gpointer*)(((gint32)(regs [reg])) + disp);
+ cinfo = get_call_info (NULL, NULL, sig, FALSE);
+
+ /*
+ * The stack looks like:
+ * <other args>
+ * <this=delegate>
+ * <possible vtype return address>
+ * <return addr>
+ * <4 pointers pushed by mono_arch_create_trampoline_code ()>
+ */
+ res = (((MonoObject**)esp) [5 + (cinfo->args [0].offset / 4)]);
+ g_free (cinfo);
+ return res;
+}
+
+#define MAX_ARCH_DELEGATE_PARAMS 10
+
+gpointer
+mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
+{
+ guint8 *code, *start;
+
+ if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
+ return NULL;
+
+ /* FIXME: Support more cases */
+ if (MONO_TYPE_ISSTRUCT (sig->ret))
+ return NULL;
+
+ /*
+ * The stack contains:
+ * <delegate>
+ * <return addr>
+ */
+
+ 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 */
+ x86_mov_reg_membase (code, X86_EAX, X86_ESP, 4, 4);
+ x86_mov_reg_membase (code, X86_ECX, X86_EAX, G_STRUCT_OFFSET (MonoDelegate, target), 4);
+ x86_mov_membase_reg (code, X86_ESP, 4, X86_ECX, 4);
+ x86_jump_membase (code, X86_EAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
+
+ g_assert ((code - start) < 64);
+
+ cached = start;
+ mono_debug_add_delegate_trampoline (start, code - start);
+ mono_mini_arch_unlock ();
+ } else {
+ static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
+ int i = 0;
+ /* 8 for mov_reg and jump, plus 8 for each parameter */
+ int code_reserve = 8 + (sig->param_count * 8);
+
+ for (i = 0; i < sig->param_count; ++i)
+ if (!mono_is_regsize_var (sig->params [i]))
+ return NULL;
+
+ mono_mini_arch_lock ();
+ code = cache [sig->param_count];
+ if (code) {
+ mono_mini_arch_unlock ();
+ return code;
+ }
+
+ /*
+ * The stack contains:
+ * <args in reverse order>
+ * <delegate>
+ * <return addr>
+ *
+ * and we need:
+ * <args in reverse order>
+ * <return addr>
+ *
+ * without unbalancing the stack.
+ * So move each arg up a spot in the stack (overwriting un-needed 'this' arg)
+ * and leaving original spot of first arg as placeholder in stack so
+ * when callee pops stack everything works.
+ */
+
+ start = code = mono_global_codeman_reserve (code_reserve);
+
+ /* store delegate for access to method_ptr */
+ x86_mov_reg_membase (code, X86_ECX, X86_ESP, 4, 4);
+
+ /* move args up */
+ for (i = 0; i < sig->param_count; ++i) {
+ x86_mov_reg_membase (code, X86_EAX, X86_ESP, (i+2)*4, 4);
+ x86_mov_membase_reg (code, X86_ESP, (i+1)*4, X86_EAX, 4);
+ }
+
+ x86_jump_membase (code, X86_ECX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
+
+ g_assert ((code - start) < code_reserve);
+
+ cache [sig->param_count] = start;
+
+ mono_debug_add_delegate_trampoline (start, code - start);
+ mono_mini_arch_unlock ();
}
- return NULL;
+ return start;
}