2008-02-08 Zoltan Varga <vargaz@gmail.com>
[mono.git] / mono / mini / mini-x86.c
index b4513f69c40e8d2832cfddc05912d6a03a89b70f..af13147e58d9226663db4ec49f564643fef9e02b 100644 (file)
 #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"
@@ -42,6 +45,11 @@ static gboolean is_win32 = TRUE;
 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
@@ -53,16 +61,19 @@ static gboolean is_win32 = FALSE;
 #define CALLCONV_IS_STDCALL(sig) (((sig)->call_convention) == MONO_CALL_STDCALL)
 #endif
 
-#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";
        }
@@ -70,8 +81,28 @@ mono_arch_regname (int reg) {
 }
 
 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 {
@@ -167,7 +198,7 @@ add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
 
 
 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)
 {
@@ -178,7 +209,7 @@ add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
        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) {
@@ -232,13 +263,14 @@ add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
  * For x86 win32, see ???.
  */
 static CallInfo*
-get_call_info (MonoMemPool *mp, 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;
 
        if (mp)
                cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
@@ -251,6 +283,7 @@ get_call_info (MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
        /* 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:
@@ -293,7 +326,7 @@ get_call_info (MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
                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);
@@ -346,6 +379,7 @@ get_call_info (MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
                        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:
@@ -379,7 +413,7 @@ get_call_info (MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
                        }
                        /* 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);
@@ -442,7 +476,7 @@ mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJit
        int offset = 8;
        CallInfo *cinfo;
 
-       cinfo = get_call_info (NULL, csig, FALSE);
+       cinfo = get_call_info (NULL, NULL, csig, FALSE);
 
        if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
                frame_size += sizeof (gpointer);
@@ -464,7 +498,7 @@ mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJit
                        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;
                }
 
@@ -592,6 +626,24 @@ mono_arch_cpu_init (void)
 #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.
  */
@@ -612,6 +664,10 @@ mono_arch_cpu_optimizazions (guint32 *exclude_mask)
                                *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;
 }
@@ -748,7 +804,7 @@ mono_arch_allocate_vars (MonoCompile *cfg)
        header = mono_method_get_header (cfg->method);
        sig = mono_method_signature (cfg->method);
 
-       cinfo = get_call_info (cfg->mempool, sig, FALSE);
+       cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
 
        cfg->frame_reg = MONO_ARCH_BASEREG;
        offset = 0;
@@ -854,7 +910,7 @@ mono_arch_create_vars (MonoCompile *cfg)
 
        sig = mono_method_signature (cfg->method);
 
-       cinfo = get_call_info (cfg->mempool, sig, FALSE);
+       cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
 
        if (cinfo->ret.storage == ArgValuetypeInReg)
                cfg->ret_var_is_local = TRUE;
@@ -891,9 +947,34 @@ emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call)
 
        arg->inst_left = sig_arg;
        arg->type = STACK_PTR;
-       /* prepend, so they get reversed */
-       arg->next = call->out_args;
-       call->out_args = arg;
+       MONO_INST_LIST_ADD (&arg->node, &call->out_args);
+}
+
+/*
+ * 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;
 }
 
 /* 
@@ -909,11 +990,12 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *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 (cfg->mempool, 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);
@@ -943,9 +1025,7 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                        arg->cil_code = in->cil_code;
                        arg->inst_left = in;
                        arg->type = in->type;
-                       /* prepend, so they get reversed */
-                       arg->next = call->out_args;
-                       call->out_args = arg;
+                       MONO_INST_LIST_ADD (&arg->node, &call->out_args);
 
                        if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(t))) {
                                guint32 size, align;
@@ -959,7 +1039,7 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                                                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;
@@ -972,11 +1052,30 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                                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:
@@ -1006,14 +1105,12 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                        zero_inst->inst_p0 = 0;
                        arg->inst_left = zero_inst;
                        arg->type = STACK_PTR;
-                       /* prepend, so they get reversed */
-                       arg->next = call->out_args;
-                       call->out_args = arg;
-               }
-               else
+                       MONO_INST_LIST_ADD (&arg->node, &call->out_args);
+               } else {
                        /* if the function returns a struct, the called method already does a ret $0x4 */
                        if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret))
                                cinfo->stack_usage -= 4;
+               }
        }
        
        call->stack_usage = cinfo->stack_usage;
@@ -1022,8 +1119,7 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
        if (cinfo->need_stack_align) {
                MONO_INST_NEW (cfg, arg, OP_X86_OUTARG_ALIGN_STACK);
                arg->inst_c0 = cinfo->stack_align_amount;
-               arg->next = call->out_args;
-               call->out_args = arg;
+               MONO_INST_LIST_ADD (&arg->node, &call->out_args);
         }
 #endif 
 
@@ -1237,17 +1333,17 @@ emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer dat
 #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:
+ * mono_peephole_pass_1:
  *
  *   Perform peephole opts which should/can be performed before local regalloc
  */
-static void
-peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
+void
+mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
 {
-       MonoInst *ins, *last_ins = NULL;
-       ins = bb->code;
+       MonoInst *ins, *n;
 
-       while (ins) {
+       MONO_INST_LIST_FOR_EACH_ENTRY_SAFE (ins, n, &bb->ins_list, node) {
+               MonoInst *last_ins = mono_inst_list_prev (&ins->node, &bb->ins_list);
                switch (ins->opcode) {
                case OP_IADD_IMM:
                case OP_ADD_IMM:
@@ -1317,8 +1413,7 @@ peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
                            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;                                
+                                       MONO_DEL_INS (ins);
                                        continue;
                                } else {
                                        //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
@@ -1343,8 +1438,7 @@ peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
                              ins->inst_offset == last_ins->inst_offset) {
 
                                if (ins->dreg == last_ins->dreg) {
-                                       last_ins->next = ins->next;                             
-                                       ins = ins->next;                                
+                                       MONO_DEL_INS (ins);
                                        continue;
                                } else {
                                        ins->opcode = OP_MOVE;
@@ -1385,7 +1479,7 @@ peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
                                (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->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? OP_ICONV_TO_I1 : OP_ICONV_TO_U1;
                                ins->sreg1 = last_ins->sreg1;
                        }
                        break;
@@ -1401,12 +1495,10 @@ peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
                        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->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? OP_ICONV_TO_I2 : OP_ICONV_TO_U2;
                                ins->sreg1 = last_ins->sreg1;
                        }
                        break;
-               case CEE_CONV_I4:
-               case CEE_CONV_U4:
                case OP_ICONV_TO_I4:
                case OP_MOVE:
                        /*
@@ -1415,9 +1507,7 @@ peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
                         * OP_MOVE reg, reg 
                         */
                        if (ins->dreg == ins->sreg1) {
-                               if (last_ins)
-                                       last_ins->next = ins->next;                             
-                               ins = ins->next;
+                               MONO_DEL_INS (ins);
                                continue;
                        }
                        /* 
@@ -1429,8 +1519,7 @@ peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
                        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;                                
+                               MONO_DEL_INS (ins);
                                continue;
                        }
                        break;
@@ -1445,25 +1534,26 @@ peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
                        }
                        break;
                }
-               last_ins = ins;
-               ins = ins->next;
        }
-       bb->last_ins = last_ins;
 }
 
-static void
-peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
+void
+mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
 {
-       MonoInst *ins, *last_ins = NULL;
-       ins = bb->code;
+       MonoInst *ins, *n;
 
-       while (ins) {
+       MONO_INST_LIST_FOR_EACH_ENTRY_SAFE (ins, n, &bb->ins_list, node) {
+               MonoInst *last_ins = mono_inst_list_prev (&ins->node, &bb->ins_list);
 
                switch (ins->opcode) {
-               case OP_ICONST:
+               case OP_ICONST: {
+                       MonoInst *next;
+
                        /* 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)))) {
+                       next = mono_inst_list_next (&ins->node, &bb->ins_list);
+                       if (ins->inst_c0 == 0 && (!next ||
+                                       (next && INST_IGNORES_CFLAGS (next->opcode)))) {
                                MonoInst *ins2;
 
                                ins->opcode = OP_IXOR;
@@ -1474,23 +1564,22 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                                 * 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) {
+                               for (ins2 = mono_inst_list_next (&ins->node, &bb->ins_list); ins2;
+                                               ins2 = mono_inst_list_next (&ins2->node, &bb->ins_list)) {
                                        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)) {
+                                       } 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)) {
+                                       } else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM)) {
                                                /* Continue iteration */
-                                       }
-                                       else
+                                       } else
                                                break;
                                }
                        }
                        break;
+               }
                case OP_IADD_IMM:
                case OP_ADD_IMM:
                        if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
@@ -1539,8 +1628,7 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                            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;                                
+                                       MONO_DEL_INS (ins);
                                        continue;
                                } else {
                                        //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
@@ -1565,8 +1653,7 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                              ins->inst_offset == last_ins->inst_offset) {
 
                                if (ins->dreg == last_ins->dreg) {
-                                       last_ins->next = ins->next;                             
-                                       ins = ins->next;                                
+                                       MONO_DEL_INS (ins);
                                        continue;
                                } else {
                                        ins->opcode = OP_MOVE;
@@ -1607,7 +1694,7 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                                (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->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? OP_ICONV_TO_I1 : OP_ICONV_TO_U1;
                                ins->sreg1 = last_ins->sreg1;
                        }
                        break;
@@ -1623,12 +1710,10 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                        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->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? OP_ICONV_TO_I2 : OP_ICONV_TO_U2;
                                ins->sreg1 = last_ins->sreg1;
                        }
                        break;
-               case CEE_CONV_I4:
-               case CEE_CONV_U4:
                case OP_ICONV_TO_I4:
                case OP_MOVE:
                        /*
@@ -1637,9 +1722,7 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                         * OP_MOVE reg, reg 
                         */
                        if (ins->dreg == ins->sreg1) {
-                               if (last_ins)
-                                       last_ins->next = ins->next;                             
-                               ins = ins->next;
+                               MONO_DEL_INS (ins);
                                continue;
                        }
                        /* 
@@ -1651,8 +1734,7 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                        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;                                
+                               MONO_DEL_INS (ins);
                                continue;
                        }
                        break;
@@ -1666,10 +1748,12 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                        }
                        break;
                }
-               last_ins = ins;
-               ins = ins->next;
        }
-       bb->last_ins = last_ins;
+}
+
+void
+mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
+{
 }
 
 static const int 
@@ -1692,18 +1776,24 @@ cc_signed_table [] = {
        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);
@@ -1848,7 +1938,7 @@ emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
 
        /* Move return value to the target register */
        switch (ins->opcode) {
-       case CEE_CALL:
+       case OP_CALL:
        case OP_CALL_REG:
        case OP_CALL_MEMBASE:
                if (ins->dreg != X86_EAX)
@@ -1857,7 +1947,7 @@ emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
        case OP_VCALL:
        case OP_VCALL_REG:
        case OP_VCALL_MEMBASE:
-               cinfo = get_call_info (cfg->mempool, ((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);
@@ -1939,7 +2029,7 @@ emit_load_volatile_arguments (MonoCompile *cfg, guint8 *code)
 
        sig = mono_method_signature (method);
 
-       cinfo = get_call_info (cfg->mempool, sig, FALSE);
+       cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
        
        /* This is the opposite of the code in emit_prolog */
 
@@ -1993,13 +2083,8 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
        MonoCallInst *call;
        guint offset;
        guint8 *code = cfg->native_code + cfg->code_len;
-       MonoInst *last_ins = NULL;
-       guint last_offset = 0;
        int max_len, cpos;
 
-       if (cfg->opt & MONO_OPT_PEEPHOLE)
-               peephole_pass (cfg, bb);
-
        if (cfg->opt & MONO_OPT_LOOP) {
                int pad, align = LOOP_ALIGNMENT;
                /* set alignment depending on cpu */
@@ -2031,20 +2116,20 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 
        mono_debug_open_block (cfg, bb, offset);
 
-       ins = bb->code;
-       while (ins) {
+       MONO_BB_FOR_EACH_INS (bb, ins) {
                offset = code - cfg->native_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:
@@ -2078,11 +2163,6 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_STOREI4_MEMBASE_REG:
                        x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
                        break;
-               case CEE_LDIND_I:
-               case CEE_LDIND_I4:
-               case CEE_LDIND_U4:
-                       x86_mov_reg_mem (code, ins->dreg, ins->inst_p0, 4);
-                       break;
                case OP_LOADU4_MEM:
                        x86_mov_reg_imm (code, ins->dreg, ins->inst_p0);
                        x86_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
@@ -2104,16 +2184,18 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_LOADI2_MEMBASE:
                        x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
                        break;
-               case CEE_CONV_I1:
+               case OP_ICONV_TO_I1:
+               case OP_SEXT_I1:
                        x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
                        break;
-               case CEE_CONV_I2:
+               case OP_ICONV_TO_I2:
+               case OP_SEXT_I2:
                        x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
                        break;
-               case CEE_CONV_U1:
+               case OP_ICONV_TO_U1:
                        x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
                        break;
-               case CEE_CONV_U2:
+               case OP_ICONV_TO_U2:
                        x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
                        break;
                case OP_COMPARE:
@@ -2180,7 +2262,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        x86_breakpoint (code);
                        break;
                case OP_ADDCC:
-               case CEE_ADD:
+               case OP_IADD:
                        x86_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
                        break;
                case OP_ADC:
@@ -2194,7 +2276,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        x86_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
                        break;
                case OP_SUBCC:
-               case CEE_SUB:
+               case OP_ISUB:
                        x86_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
                        break;
                case OP_SBB:
@@ -2207,56 +2289,68 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_SBB_IMM:
                        x86_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
                        break;
-               case CEE_AND:
+               case OP_IAND:
                        x86_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
                        break;
                case OP_AND_IMM:
                        x86_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
                        break;
-               case CEE_DIV:
-                       x86_cdq (code);
-                       x86_div_reg (code, ins->sreg2, TRUE);
+               case OP_IDIV:
+               case OP_IREM:
+                       /* 
+                        * The code is the same for div/rem, the allocator will allocate dreg
+                        * to RAX/RDX as appropriate.
+                        */
+                       if (ins->sreg2 == X86_EDX) {
+                               /* cdq clobbers this */
+                               x86_push_reg (code, ins->sreg2);
+                               x86_cdq (code);
+                               x86_div_membase (code, X86_ESP, 0, TRUE);
+                               x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);                            
+                       } else {
+                               x86_cdq (code);
+                               x86_div_reg (code, ins->sreg2, TRUE);
+                       }
                        break;
-               case CEE_DIV_UN:
-                       x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
-                       x86_div_reg (code, ins->sreg2, FALSE);
+               case OP_IDIV_UN:
+               case OP_IREM_UN:
+                       if (ins->sreg2 == X86_EDX) {
+                               x86_push_reg (code, ins->sreg2);
+                               x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
+                               x86_div_membase (code, X86_ESP, 0, FALSE);
+                               x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);                            
+                       } else {
+                               x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
+                               x86_div_reg (code, ins->sreg2, FALSE);
+                       }
                        break;
                case OP_DIV_IMM:
                        x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
                        x86_cdq (code);
                        x86_div_reg (code, ins->sreg2, TRUE);
                        break;
-               case CEE_REM:
-                       x86_cdq (code);
-                       x86_div_reg (code, ins->sreg2, TRUE);
-                       break;
-               case CEE_REM_UN:
-                       x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
-                       x86_div_reg (code, ins->sreg2, FALSE);
-                       break;
                case OP_REM_IMM:
                        x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
                        x86_cdq (code);
                        x86_div_reg (code, ins->sreg2, TRUE);
                        break;
-               case CEE_OR:
+               case OP_IOR:
                        x86_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
                        break;
                case OP_OR_IMM:
                        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:
                        x86_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
                        break;
-               case CEE_SHL:
+               case OP_ISHL:
                        g_assert (ins->sreg2 == X86_ECX);
                        x86_shift_reg (code, X86_SHL, ins->dreg);
                        break;
-               case CEE_SHR:
+               case OP_ISHR:
                        g_assert (ins->sreg2 == X86_ECX);
                        x86_shift_reg (code, X86_SAR, ins->dreg);
                        break;
@@ -2266,7 +2360,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_SHR_UN_IMM:
                        x86_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
                        break;
-               case CEE_SHR_UN:
+               case OP_ISHR_UN:
                        g_assert (ins->sreg2 == X86_ECX);
                        x86_shift_reg (code, X86_SHR, ins->dreg);
                        break;
@@ -2354,19 +2448,14 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                x86_shift_reg_imm (code, X86_SHR, ins->backend.reg3, ins->inst_imm);
                        }
                        break;
-               case CEE_NOT:
+               case OP_INOT:
                        x86_not_reg (code, ins->sreg1);
                        break;
-               case CEE_NEG:
+               case OP_INEG:
                        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:
+
+               case OP_IMUL:
                        x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
                        break;
                case OP_MUL_IMM:
@@ -2427,11 +2516,11 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                break;
                        }
                        break;
-               case CEE_MUL_OVF:
+               case OP_IMUL_OVF:
                        x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
                        EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
                        break;
-               case CEE_MUL_OVF_UN: {
+               case OP_IMUL_OVF_UN: {
                        /* the mul operation and the exception check should most likely be split */
                        int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
                        /*g_assert (ins->sreg2 == X86_EAX);
@@ -2495,12 +2584,9 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
                        x86_push_membase (code, ins->inst_basereg, 0xf0f0f0f0);
                        break;
-               case CEE_CONV_I4:
                case OP_MOVE:
                        x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
                        break;
-               case CEE_CONV_U4:
-                       g_assert_not_reached ();
                case OP_JMP: {
                        /*
                         * Note: this 'frame destruction' logic is useful for tail calls, too.
@@ -2560,7 +2646,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_LCALL:
                case OP_VCALL:
                case OP_VOIDCALL:
-               case CEE_CALL:
+               case OP_CALL:
                        call = (MonoCallInst*)ins;
                        if (ins->flags & MONO_INST_HAS_METHOD)
                                code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
@@ -2664,9 +2750,6 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        code = mono_emit_stack_alloc (code, ins);
                        x86_mov_reg_reg (code, ins->dreg, X86_ESP, 4);
                        break;
-               case CEE_RET:
-                       x86_ret (code);
-                       break;
                case OP_THROW: {
                        x86_push_reg (code, ins->sreg1);
                        code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, 
@@ -2690,13 +2773,29 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        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;
                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;
                        if (ins->flags & MONO_INST_BRLABEL) {
                                if (ins->inst_i0->inst_c0) {
                                        x86_jump_code (code, cfg->native_code + ins->inst_i0->inst_c0);
@@ -2751,16 +2850,16 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_COND_EXC_NC:
                        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 CEE_BEQ:
-               case CEE_BNE_UN:
-               case CEE_BLT:
-               case CEE_BLT_UN:
-               case CEE_BGT:
-               case CEE_BGT_UN:
-               case CEE_BGE:
-               case CEE_BGE_UN:
-               case CEE_BLE:
-               case CEE_BLE_UN:
+               case OP_IBEQ:
+               case OP_IBNE_UN:
+               case OP_IBLT:
+               case OP_IBLT_UN:
+               case OP_IBGT:
+               case OP_IBGT_UN:
+               case OP_IBGE:
+               case OP_IBGE_UN:
+               case OP_IBLE:
+               case OP_IBLE_UN:
                        EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
                        break;
 
@@ -2824,8 +2923,8 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_LOADR4_MEMBASE:
                        x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
                        break;
-               case CEE_CONV_R4: /* FIXME: change precision */
-               case CEE_CONV_R8:
+               case OP_ICONV_TO_R4: /* FIXME: change precision */
+               case OP_ICONV_TO_R8:
                        x86_push_reg (code, ins->sreg1);
                        x86_fild_membase (code, X86_ESP, 0, FALSE);
                        x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
@@ -3297,13 +3396,22 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
                        break;
                case OP_CKFINITE: {
+                       guchar *br1;
                        x86_push_reg (code, X86_EAX);
                        x86_fxam (code);
                        x86_fnstsw (code);
                        x86_alu_reg_imm (code, X86_AND, X86_EAX, 0x4100);
                        x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
                        x86_pop_reg (code, X86_EAX);
+
+                       /* Have to clean up the fp stack before throwing the exception */
+                       br1 = code;
+                       x86_branch8 (code, X86_CC_NE, 0, FALSE);
+
+                       x86_fstp (code, 0);                     
                        EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
+
+                       x86_patch (br1, code);
                        break;
                }
                case OP_TLS_GET: {
@@ -3424,22 +3532,17 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        break;
                }
                default:
-                       g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
+                       g_warning ("unknown opcode %s\n", mono_inst_name (ins->opcode));
                        g_assert_not_reached ();
                }
 
-               if ((code - cfg->native_code - offset) > max_len) {
+               if (G_UNLIKELY ((code - cfg->native_code - offset) > max_len)) {
                        g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %d)",
                                   mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
                        g_assert_not_reached ();
                }
               
                cpos += max_len;
-
-               last_ins = ins;
-               last_offset = offset;
-               
-               ins = ins->next;
        }
 
        cfg->code_len = code - cfg->native_code;
@@ -3513,7 +3616,11 @@ mono_arch_emit_prolog (MonoCompile *cfg)
        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);
@@ -3523,14 +3630,19 @@ mono_arch_emit_prolog (MonoCompile *cfg)
        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);
@@ -3569,8 +3681,8 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                        /* %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 */
@@ -3595,8 +3707,8 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                                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); 
@@ -3668,7 +3780,7 @@ mono_arch_emit_prolog (MonoCompile *cfg)
        max_offset = 0;
        if (cfg->opt & MONO_OPT_BRANCH) {
                for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
-                       MonoInst *ins = bb->code;
+                       MonoInst *ins;
                        bb->max_offset = max_offset;
 
                        if (cfg->prof_options & MONO_PROFILE_COVERAGE)
@@ -3677,12 +3789,11 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                        if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
                                max_offset += LOOP_ALIGNMENT;
 
-                       while (ins) {
+                       MONO_BB_FOR_EACH_INS (bb, ins) {
                                if (ins->opcode == OP_LABEL)
                                        ins->inst_c1 = max_offset;
                                
                                max_offset += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
-                               ins = ins->next;
                        }
                }
        }
@@ -3722,9 +3833,6 @@ mono_arch_emit_epilog (MonoCompile *cfg)
        
        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;
@@ -3757,7 +3865,7 @@ mono_arch_emit_epilog (MonoCompile *cfg)
                        x86_mov_mem_reg (code, lmf_tls_offset, X86_ECX, 4);
                } else {
                        /* Find a spare register */
-                       switch (sig->ret->type) {
+                       switch (mono_type_get_underlying_type (sig->ret)->type) {
                        case MONO_TYPE_I8:
                        case MONO_TYPE_U8:
                                prev_lmf_reg = X86_EDI;
@@ -3817,7 +3925,7 @@ mono_arch_emit_epilog (MonoCompile *cfg)
        }
 
        /* Load returned vtypes into registers if needed */
-       cinfo = get_call_info (cfg->mempool, 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]) {
@@ -4027,7 +4135,7 @@ void
 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 (cfg->mempool, inst->signature, FALSE);
+       CallInfo *cinfo = get_call_info (cfg, cfg->mempool, inst->signature, FALSE);
 
        /* add the this argument */
        if (this_reg != -1) {
@@ -4083,6 +4191,137 @@ mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_re
        }
 }
 
+#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)
 {
@@ -4116,42 +4355,6 @@ mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethod
                        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;
@@ -4219,12 +4422,40 @@ mono_arch_get_patch_offset (guint8 *code)
        }
 }
 
-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
@@ -4247,6 +4478,15 @@ mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
                 */
                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];
@@ -4269,25 +4509,137 @@ mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
                        return NULL;
        }
 
-       return (gpointer*)(((gint32)(regs [reg])) + disp);
+       *displacement = disp;
+       return regs [reg];
+}
+
+gpointer*
+mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
+{
+       gpointer vt;
+       int displacement;
+       vt = mono_arch_get_vcall_slot (code, regs, &displacement);
+       if (!vt)
+               return NULL;
+       return (gpointer*)((char*)vt + displacement);
 }
 
-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)
 {
-       guint8 reg = 0;
-       gint32 disp = 0;
+       guint32 esp = regs [X86_ESP];
+       CallInfo *cinfo;
+       gpointer res;
 
-       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];
+       cinfo = get_call_info (NULL, NULL, sig, FALSE);
 
-               if (reg == X86_EAX)
-                       return NULL;
-               else
-                       return (gpointer*)(((gint32)(regs [reg])) + disp);
+       /*
+        * 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;
 }