ImageButtonTest.cs added new tests for 2.0 and new resource NoEventValidation.aspx
[mono.git] / mono / mini / mini-ia64.c
index 200b19586c08ef545aa80ba009a722b223c84380..4d67083dc11b09f640149cec68072701143b0716 100644 (file)
@@ -54,13 +54,12 @@ static const char*const * ins_spec = ia64_desc;
  * ia64_codegen_set_one_ins_per_bundle () at those places.
  */
 
-#define SIGNAL_STACK_SIZE (64 * 1024)
-
 #define ARGS_OFFSET 16
 
 #define GP_SCRATCH_REG 31
 #define GP_SCRATCH_REG2 30
 #define FP_SCRATCH_REG 32
+#define FP_SCRATCH_REG2 33
 
 #define LOOP_ALIGNMENT 8
 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
@@ -400,6 +399,13 @@ get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
                        cinfo->ret.storage = ArgInFloatReg;
                        cinfo->ret.reg = 8;
                        break;
+               case MONO_TYPE_GENERICINST:
+                       if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
+                               cinfo->ret.storage = ArgInIReg;
+                               cinfo->ret.reg = IA64_R8;
+                               break;
+                       }
+                       /* Fall through */
                case MONO_TYPE_VALUETYPE:
                case MONO_TYPE_TYPEDBYREF: {
                        guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
@@ -480,6 +486,12 @@ get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
                case MONO_TYPE_ARRAY:
                        add_general (&gr, &stack_size, ainfo);
                        break;
+               case MONO_TYPE_GENERICINST:
+                       if (!mono_type_generic_inst_is_valuetype (sig->params [i])) {
+                               add_general (&gr, &stack_size, ainfo);
+                               break;
+                       }
+                       /* Fall through */
                case MONO_TYPE_VALUETYPE:
                case MONO_TYPE_TYPEDBYREF:
                        /* FIXME: */
@@ -597,6 +609,10 @@ is_regsize_var (MonoType *t) {
        case MONO_TYPE_SZARRAY:
        case MONO_TYPE_ARRAY:
                return TRUE;
+       case MONO_TYPE_GENERICINST:
+               if (!mono_type_generic_inst_is_valuetype (t))
+                       return TRUE;
+               return FALSE;
        case MONO_TYPE_VALUETYPE:
                return FALSE;
        }
@@ -661,7 +677,8 @@ static void
 mono_ia64_alloc_stacked_registers (MonoCompile *cfg)
 {
        CallInfo *cinfo;
-       guint32 reserved_regs = 3;
+       guint32 reserved_regs;
+       MonoMethodHeader *header;
 
        if (cfg->arch.reg_local0 > 0)
                /* Already done */
@@ -669,8 +686,10 @@ mono_ia64_alloc_stacked_registers (MonoCompile *cfg)
 
        cinfo = get_call_info (mono_method_signature (cfg->method), FALSE);
 
-       /* Three registers are reserved for use by the prolog/epilog */
-       reserved_regs = 3;
+       header = mono_method_get_header (cfg->method);
+       
+       /* Some registers are reserved for use by the prolog/epilog */
+       reserved_regs = header->num_clauses ? 4 : 3;
 
        if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
                (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
@@ -687,11 +706,20 @@ mono_ia64_alloc_stacked_registers (MonoCompile *cfg)
 
        cfg->arch.reg_saved_ar_pfs = cfg->arch.reg_local0 - 1;
        cfg->arch.reg_saved_b0 = cfg->arch.reg_local0 - 2;
-       cfg->arch.reg_saved_sp = cfg->arch.reg_local0 - 3;
+       cfg->arch.reg_fp = cfg->arch.reg_local0 - 3;
+
+       /* 
+        * Frames without handlers save sp to fp, frames with handlers save it into
+        * a dedicated register.
+        */
+       if (header->num_clauses)
+               cfg->arch.reg_saved_sp = cfg->arch.reg_local0 - 4;
+       else
+               cfg->arch.reg_saved_sp = cfg->arch.reg_fp;
 
        if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
                (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
-               cfg->arch.reg_saved_return_val = cfg->arch.reg_local0 - 4;
+               cfg->arch.reg_saved_return_val = cfg->arch.reg_local0 - reserved_regs;
        }
 
        /* 
@@ -792,7 +820,7 @@ mono_arch_allocate_vars (MonoCompile *cfg)
        }
        else {
                /* Locals are allocated backwards from %fp */
-               cfg->frame_reg = cfg->arch.reg_saved_sp;
+               cfg->frame_reg = cfg->arch.reg_fp;
                offset = 0;
        }
 
@@ -957,14 +985,14 @@ add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, MonoInst *arg, ArgStorage
                arg->opcode = OP_OUTARG_REG;
                arg->inst_left = tree;
                arg->inst_right = (MonoInst*)call;
-               arg->unused = reg;
+               arg->backend.reg3 = reg;
                call->used_iregs |= 1 << reg;
                break;
        case ArgInFloatReg:
                arg->opcode = OP_OUTARG_FREG;
                arg->inst_left = tree;
                arg->inst_right = (MonoInst*)call;
-               arg->unused = reg;
+               arg->backend.reg3 = reg;
                call->used_fregs |= 1 << reg;
                break;
        default:
@@ -972,12 +1000,47 @@ add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, MonoInst *arg, ArgStorage
        }
 }
 
+static void
+emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
+{
+       MonoInst *arg;
+       MonoMethodSignature *tmp_sig;
+       MonoInst *sig_arg;
+
+       /* FIXME: Add support for signature tokens to AOT */
+       cfg->disable_aot = TRUE;
+
+       g_assert (cinfo->sig_cookie.storage == ArgOnStack);
+
+       /*
+        * mono_ArgIterator_Setup assumes the signature cookie is 
+        * passed first and all the arguments which were before it are
+        * passed on the stack after the signature. So compensate by 
+        * passing a different signature.
+        */
+       tmp_sig = mono_metadata_signature_dup (call->signature);
+       tmp_sig->param_count -= call->signature->sentinelpos;
+       tmp_sig->sentinelpos = 0;
+       memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
+
+       MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
+       sig_arg->inst_p0 = tmp_sig;
+
+       MONO_INST_NEW (cfg, arg, OP_OUTARG);
+       arg->inst_left = sig_arg;
+       arg->inst_imm = 16 + cinfo->sig_cookie.offset;
+       arg->type = STACK_PTR;
+
+       /* prepend, so they get reversed */
+       arg->next = call->out_args;
+       call->out_args = arg;
+}
+
 /* 
  * take the arguments and generate the arch-specific
  * instructions to properly call the function in call.
  * This includes pushing, moving arguments to the right register
  * etc.
- * Issue: who does the spilling if needed, and when?
  */
 MonoCallInst*
 mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual)
@@ -1007,37 +1070,8 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                ainfo = cinfo->args + i;
 
                if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
-                       MonoMethodSignature *tmp_sig;
-
                        /* Emit the signature cookie just before the implicit arguments */
-                       MonoInst *sig_arg;
-                       /* FIXME: Add support for signature tokens to AOT */
-                       cfg->disable_aot = TRUE;
-
-                       g_assert (cinfo->sig_cookie.storage == ArgOnStack);
-
-                       /*
-                        * mono_ArgIterator_Setup assumes the signature cookie is 
-                        * passed first and all the arguments which were before it are
-                        * passed on the stack after the signature. So compensate by 
-                        * passing a different signature.
-                        */
-                       tmp_sig = mono_metadata_signature_dup (call->signature);
-                       tmp_sig->param_count -= call->signature->sentinelpos;
-                       tmp_sig->sentinelpos = 0;
-                       memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
-
-                       MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
-                       sig_arg->inst_p0 = tmp_sig;
-
-                       MONO_INST_NEW (cfg, arg, OP_OUTARG);
-                       arg->inst_left = sig_arg;
-                       arg->inst_imm = 16 + cinfo->sig_cookie.offset;
-                       arg->type = STACK_PTR;
-
-                       /* prepend, so they get reversed */
-                       arg->next = call->out_args;
-                       call->out_args = arg;
+                       emit_sig_cookie (cfg, call, cinfo);
                }
 
                if (is_virtual && i == 0) {
@@ -1062,7 +1096,7 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
 
                        if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(arg_type))) {
                                MonoInst *stack_addr;
-                               gint align;
+                               guint32 align;
                                guint32 size;
 
                                if (arg_type->type == MONO_TYPE_TYPEDBYREF) {
@@ -1084,32 +1118,32 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
 
                                if (ainfo->storage == ArgAggregate) {
                                        MonoInst *vtaddr, *load, *load2, *offset_ins, *set_reg;
-                                       int slot;
+                                       int slot, j;
 
                                        vtaddr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
 
                                        /* 
                                         * Part of the structure is passed in registers.
                                         */
-                                       for (i = 0; i < ainfo->nregs; ++i) {
+                                       for (j = 0; j < ainfo->nregs; ++j) {
                                                int offset, load_op, dest_reg, arg_storage;
 
-                                               slot = ainfo->reg + i;
+                                               slot = ainfo->reg + j;
                                                
                                                if (ainfo->atype == AggregateSingleHFA) {
                                                        load_op = CEE_LDIND_R4;
-                                                       offset = i * 4;
-                                                       dest_reg = ainfo->reg + i;
+                                                       offset = j * 4;
+                                                       dest_reg = ainfo->reg + j;
                                                        arg_storage = ArgInFloatReg;
                                                } else if (ainfo->atype == AggregateDoubleHFA) {
                                                        load_op = CEE_LDIND_R8;
-                                                       offset = i * 8;
-                                                       dest_reg = ainfo->reg + i;
+                                                       offset = j * 8;
+                                                       dest_reg = ainfo->reg + j;
                                                        arg_storage = ArgInFloatReg;
                                                } else {
                                                        load_op = CEE_LDIND_I;
-                                                       offset = i * 8;
-                                                       dest_reg = cfg->arch.reg_out0 + ainfo->reg + i;
+                                                       offset = j * 8;
+                                                       dest_reg = cfg->arch.reg_out0 + ainfo->reg + j;
                                                        arg_storage = ArgInIReg;
                                                }
 
@@ -1125,7 +1159,7 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                                                MONO_INST_NEW (cfg, load, load_op);
                                                load->inst_left = load2;
 
-                                               if (i == 0)
+                                               if (j == 0)
                                                        set_reg = arg;
                                                else
                                                        MONO_INST_NEW (cfg, set_reg, OP_OUTARG_REG);
@@ -1139,16 +1173,16 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                                        /* 
                                         * Part of the structure is passed on the stack.
                                         */
-                                       for (i = ainfo->nregs; i < ainfo->nslots; ++i) {
+                                       for (j = ainfo->nregs; j < ainfo->nslots; ++j) {
                                                MonoInst *outarg;
 
-                                               slot = ainfo->reg + i;
+                                               slot = ainfo->reg + j;
 
                                                MONO_INST_NEW (cfg, load, CEE_LDIND_I);
                                                load->ssa_op = MONO_SSA_LOAD;
                                                load->inst_i0 = (cfg)->varinfo [vtaddr->inst_c0];
 
-                                               NEW_ICONST (cfg, offset_ins, (i * sizeof (gpointer)));
+                                               NEW_ICONST (cfg, offset_ins, (j * sizeof (gpointer)));
                                                MONO_INST_NEW (cfg, load2, CEE_ADD);
                                                load2->inst_left = load;
                                                load2->inst_right = offset_ins;
@@ -1156,7 +1190,7 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                                                MONO_INST_NEW (cfg, load, CEE_LDIND_I);
                                                load->inst_left = load2;
 
-                                               if (i == 0)
+                                               if (j == 0)
                                                        outarg = arg;
                                                else
                                                        MONO_INST_NEW (cfg, outarg, OP_OUTARG);
@@ -1214,6 +1248,11 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
                }
        }
 
+       /* Handle the case where there are no implicit arguments */
+       if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos)) {
+               emit_sig_cookie (cfg, call, cinfo);
+       }
+
        call->stack_usage = cinfo->stack_usage;
        cfg->param_area = MAX (cfg->param_area, call->stack_usage);
        cfg->arch.n_out_regs = MAX (cfg->arch.n_out_regs, cinfo->reg_usage);
@@ -1487,7 +1526,9 @@ mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_STOREI8_MEMBASE_IMM:
                case OP_STORE_MEMBASE_IMM:
                        /* There are no store_membase instructions on ia64 */
-                       if (ia64_is_imm14 (ins->inst_offset)) {
+                       if (ins->inst_offset == 0) {
+                               temp2 = NULL;
+                       } else if (ia64_is_imm14 (ins->inst_offset)) {
                                NEW_INS (cfg, temp2, OP_ADD_IMM);
                                temp2->sreg1 = ins->inst_destbasereg;
                                temp2->inst_imm = ins->inst_offset;
@@ -1531,7 +1572,8 @@ mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                        }
 
                        ins->inst_offset = 0;
-                       ins->inst_destbasereg = temp2->dreg;
+                       if (temp2)
+                               ins->inst_destbasereg = temp2->dreg;
                        break;
                case OP_STOREI1_MEMBASE_REG:
                case OP_STOREI2_MEMBASE_REG:
@@ -1573,7 +1615,13 @@ mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_LOAD_MEMBASE:
                case OP_LOADR4_MEMBASE:
                case OP_LOADR8_MEMBASE:
-                       /* There are no load_membase instructions on ia64 */
+               case OP_ATOMIC_EXCHANGE_I4:
+               case OP_ATOMIC_EXCHANGE_I8:
+               case OP_ATOMIC_ADD_NEW_I4:
+               case OP_ATOMIC_ADD_NEW_I8:
+               case OP_ATOMIC_ADD_IMM_NEW_I4:
+               case OP_ATOMIC_ADD_IMM_NEW_I8:
+                       /* There are no membase instructions on ia64 */
                        if (ins->inst_offset == 0) {
                                break;
                        }
@@ -1889,10 +1937,11 @@ mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                        break;
                }
                case OP_MUL_IMM:
+               case OP_LMUL_IMM:
                case OP_IMUL_IMM: {
                        int i, sum_reg;
                        gboolean found = FALSE;
-                       int shl_op = ins->opcode == OP_MUL_IMM ? OP_SHL_IMM : OP_ISHL_IMM;
+                       int shl_op = ins->opcode == OP_IMUL_IMM ? OP_ISHL_IMM : OP_SHL_IMM;
 
                        /* First the easy cases */
                        if (ins->inst_imm == 1) {
@@ -2127,9 +2176,11 @@ emit_move_return_value (MonoCompile *cfg, MonoInst *ins, Ia64CodegenState code)
                break;
        case OP_VCALL:
        case OP_VCALL_REG:
-       case OP_VCALL_MEMBASE:
+       case OP_VCALL_MEMBASE: {
+               ArgStorage storage;
+
                cinfo = get_call_info (((MonoCallInst*)ins)->signature, FALSE);
-               ArgStorage storage = cinfo->ret.storage;
+               storage = cinfo->ret.storage;
 
                if (storage == ArgAggregate) {
                        MonoInst *local = (MonoInst*)cfg->arch.ret_var_addr_local;
@@ -2157,6 +2208,7 @@ emit_move_return_value (MonoCompile *cfg, MonoInst *ins, Ia64CodegenState code)
                }
                g_free (cinfo);
                break;
+       }
        default:
                g_assert_not_reached ();
        }
@@ -2168,6 +2220,15 @@ emit_move_return_value (MonoCompile *cfg, MonoInst *ins, Ia64CodegenState code)
        mono_add_patch_info (cfg, code.buf + code.nins - cfg->native_code, patch_type, data); \
 } while (0)
 
+#define emit_cond_system_exception(cfg,code,exc_name,predicate) do { \
+       MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
+    if (tins == NULL) \
+        add_patch_info (cfg, code, MONO_PATCH_INFO_EXC, exc_name); \
+    else \
+               add_patch_info (cfg, code, MONO_PATCH_INFO_BB, tins->inst_true_bb); \
+       ia64_br_cond_pred (code, (predicate), 0); \
+} while (0)
+
 static Ia64CodegenState
 emit_call (MonoCompile *cfg, Ia64CodegenState code, guint32 patch_type, gconstpointer data)
 {
@@ -2175,6 +2236,7 @@ emit_call (MonoCompile *cfg, Ia64CodegenState code, guint32 patch_type, gconstpo
 
        if ((patch_type == MONO_PATCH_INFO_ABS) || (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD)) {
                /* Indirect call */
+               /* mono_arch_patch_callsite will patch this */
                /* mono_arch_nullify_class_init_trampoline will patch this */
                ia64_movl (code, GP_SCRATCH_REG, 0);
                ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
@@ -2443,6 +2505,13 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_LSHR_UN_IMM:
                        ia64_shr_u_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
                        break;
+               case CEE_MUL:
+                       /* Based on gcc code */
+                       ia64_setf_sig (code, FP_SCRATCH_REG, ins->sreg1);
+                       ia64_setf_sig (code, FP_SCRATCH_REG2, ins->sreg2);
+                       ia64_xmpy_l (code, FP_SCRATCH_REG, FP_SCRATCH_REG, FP_SCRATCH_REG2);
+                       ia64_getf_sig (code, ins->dreg, FP_SCRATCH_REG);
+                       break;
 
                case OP_STOREI1_MEMBASE_REG:
                        ia64_st1_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
@@ -2687,20 +2756,14 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 
                case OP_COND_EXC_IOV:
                case OP_COND_EXC_OV:
-                       add_patch_info (cfg, code,
-                                                                MONO_PATCH_INFO_EXC, "OverflowException");
-                       ia64_br_cond_pred (code, 6, 0);
+                       emit_cond_system_exception (cfg, code, "OverflowException", 6);
                        break;
                case OP_COND_EXC_IC:
                case OP_COND_EXC_C:
-                       add_patch_info (cfg, code,
-                                                                MONO_PATCH_INFO_EXC, "OverflowException");
-                       ia64_br_cond_pred (code, 7, 0);
+                       emit_cond_system_exception (cfg, code, "OverflowException", 7);
                        break;
                case OP_IA64_COND_EXC:
-                       add_patch_info (cfg, code,
-                                                                MONO_PATCH_INFO_EXC, ins->inst_p1);
-                       ia64_br_cond_pred (code, 6, 0);
+                       emit_cond_system_exception (cfg, code, ins->inst_p1, 6);
                        break;
                case OP_IA64_CSET:
                        ia64_mov_pred (code, 7, ins->dreg, IA64_R0);
@@ -2779,13 +2842,15 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        ia64_stfd_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
                        break;
                case OP_STORER4_MEMBASE_REG:
-                       ia64_stfs_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
+                       ia64_fnorm_s_sf (code, FP_SCRATCH_REG, ins->sreg1, 0);
+                       ia64_stfs_hint (code, ins->inst_destbasereg, FP_SCRATCH_REG, 0);
                        break;
                case OP_LOADR8_MEMBASE:
                        ia64_ldfd (code, ins->dreg, ins->inst_basereg);
                        break;
                case OP_LOADR4_MEMBASE:
                        ia64_ldfs (code, ins->dreg, ins->inst_basereg);
+                       ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
                        break;
                case CEE_CONV_R4:
                        ia64_setf_sig (code, ins->dreg, ins->sreg1);
@@ -2831,24 +2896,16 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case CEE_CKFINITE:
                        /* Quiet NaN */
                        ia64_fclass_m (code, 6, 7, ins->sreg1, 0x080);
-                       add_patch_info (cfg, code,
-                                                                MONO_PATCH_INFO_EXC, "ArithmeticException");
-                       ia64_br_cond_pred (code, 6, 0);
+                       emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
                        /* Signaling NaN */
                        ia64_fclass_m (code, 6, 7, ins->sreg1, 0x040);
-                       add_patch_info (cfg, code,
-                                                                MONO_PATCH_INFO_EXC, "ArithmeticException");
-                       ia64_br_cond_pred (code, 6, 0);
+                       emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
                        /* Positive infinity */
                        ia64_fclass_m (code, 6, 7, ins->sreg1, 0x021);
-                       add_patch_info (cfg, code,
-                                                                MONO_PATCH_INFO_EXC, "ArithmeticException");
-                       ia64_br_cond_pred (code, 6, 0);
+                       emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
                        /* Negative infinity */
                        ia64_fclass_m (code, 6, 7, ins->sreg1, 0x022);
-                       add_patch_info (cfg, code,
-                                                                MONO_PATCH_INFO_EXC, "ArithmeticException");
-                       ia64_br_cond_pred (code, 6, 0);
+                       emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
                        break;
 
                /* Calls */
@@ -2968,10 +3025,15 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 
                        break;
                }
+               case CEE_BREAK:
+                       code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, mono_arch_break);
+                       break;
 
                case OP_LOCALLOC: {
                        gint32 abi_offset;
 
+                       /* FIXME: Sigaltstack support */
+
                        /* keep alignment */
                        ia64_adds_imm (code, GP_SCRATCH_REG, MONO_ARCH_FRAME_ALIGNMENT - 1, ins->sreg1);
                        ia64_movl (code, GP_SCRATCH_REG2, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
@@ -3013,6 +3075,68 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        ia64_ld8 (code, ins->dreg, ins->dreg);
                        break;
 
+                       /* Synchronization */
+               case OP_MEMORY_BARRIER:
+                       ia64_mf (code);
+                       break;
+               case OP_ATOMIC_ADD_IMM_NEW_I4:
+                       g_assert (ins->inst_offset == 0);
+                       ia64_fetchadd4_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
+                       ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
+                       break;
+               case OP_ATOMIC_ADD_IMM_NEW_I8:
+                       g_assert (ins->inst_offset == 0);
+                       ia64_fetchadd8_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
+                       ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
+                       break;
+               case OP_ATOMIC_EXCHANGE_I4:
+                       ia64_xchg4_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
+                       ia64_sxt4 (code, ins->dreg, ins->dreg);
+                       break;
+               case OP_ATOMIC_EXCHANGE_I8:
+                       ia64_xchg8_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
+                       break;
+               case OP_ATOMIC_ADD_NEW_I4: {
+                       guint8 *label, *buf;
+
+                       /* From libatomic_ops */
+                       ia64_mf (code);
+
+                       ia64_begin_bundle (code);
+                       label = code.buf + code.nins;
+                       ia64_ld4_acq (code, GP_SCRATCH_REG, ins->sreg1);
+                       ia64_add (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, ins->sreg2);
+                       ia64_mov_to_ar_m (code, IA64_CCV, GP_SCRATCH_REG);
+                       ia64_cmpxchg4_acq_hint (code, GP_SCRATCH_REG2, ins->sreg1, GP_SCRATCH_REG2, 0);
+                       ia64_cmp4_eq (code, 6, 7, GP_SCRATCH_REG, GP_SCRATCH_REG2);
+                       buf = code.buf + code.nins;
+                       ia64_br_cond_pred (code, 7, 0);
+                       ia64_begin_bundle (code);
+                       ia64_patch (buf, label);
+                       ia64_add (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
+                       break;
+               }
+               case OP_ATOMIC_ADD_NEW_I8: {
+                       guint8 *label, *buf;
+
+                       /* From libatomic_ops */
+                       ia64_mf (code);
+
+                       ia64_begin_bundle (code);
+                       label = code.buf + code.nins;
+                       ia64_ld8_acq (code, GP_SCRATCH_REG, ins->sreg1);
+                       ia64_add (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, ins->sreg2);
+                       ia64_mov_to_ar_m (code, IA64_CCV, GP_SCRATCH_REG);
+                       ia64_cmpxchg8_acq_hint (code, GP_SCRATCH_REG2, ins->sreg1, GP_SCRATCH_REG2, 0);
+                       ia64_cmp_eq (code, 6, 7, GP_SCRATCH_REG, GP_SCRATCH_REG2);
+                       buf = code.buf + code.nins;
+                       ia64_br_cond_pred (code, 7, 0);
+                       ia64_begin_bundle (code);
+                       ia64_patch (buf, label);
+                       ia64_add (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
+                       break;
+               }
+
                        /* Exception handling */
                case OP_CALL_HANDLER:
                        /*
@@ -3038,19 +3162,32 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
 
                        /* 
-                        * We might be called by call_filter, in which case the
-                        * the register stack is not set up correctly. So do it now.
-                        * Allocate a stack frame and set the fp register from the value 
-                        * passed in by the caller.
-                        * R15 is used since it is writable using libunwind.
+                        * R15 determines our caller. It is used since it is writable using
+                        * libunwind.
                         * R15 == 0 means we are called by OP_CALL_HANDLER or via resume_context ()
+                        * R15 != 0 means we are called by call_filter ().
                         */
                        ia64_codegen_set_one_ins_per_bundle (code, TRUE);
                        ia64_cmp_eq (code, 6, 7, IA64_R15, IA64_R0);
-                       /* Alloc is not predictable so we have to use a branch */
-                       ia64_br_cond_pred (code, 6, 3);
+
+                       ia64_br_cond_pred (code, 6, 6);
+
+                       /*
+                        * Called by call_filter:
+                        * Allocate a new stack frame, and set the fp register from the 
+                        * value passed in by the caller.
+                        * We allocate a similar frame as is done by the prolog, so
+                        * if an exception is thrown while executing the filter, the
+                        * unwinder can unwind through the filter frame using the unwind
+                        * info for the prolog. 
+                        */
                        ia64_alloc (code, cfg->arch.reg_saved_ar_pfs, cfg->arch.reg_local0 - cfg->arch.reg_in0, cfg->arch.reg_out0 - cfg->arch.reg_local0, cfg->arch.n_out_regs, 0);
+                       ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
+                       ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
                        ia64_mov (code, cfg->frame_reg, IA64_R15);
+                       /* Signal to endfilter that we are called by call_filter */
+                       ia64_mov (code, GP_SCRATCH_REG, IA64_R0);
+
                        /* Save the return address */
                        ia64_adds_imm (code, GP_SCRATCH_REG2, spvar->inst_offset, cfg->frame_reg);
                        ia64_st8_hint (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 0);
@@ -3058,23 +3195,26 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 
                        break;
                }
-               case CEE_ENDFINALLY: {
-                       MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
-                       /* Return the saved arp_pfs value to call_filter */
-                       ia64_mov (code, IA64_R9, cfg->arch.reg_saved_ar_pfs);
-                       ia64_adds_imm (code, GP_SCRATCH_REG, spvar->inst_offset, cfg->frame_reg);
-                       ia64_ld8_hint (code, GP_SCRATCH_REG, GP_SCRATCH_REG, 0);
-                       ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
-                       ia64_br_cond_reg (code, IA64_B6);
-                       break;
-               }
+               case CEE_ENDFINALLY:
                case OP_ENDFILTER: {
-                       /* FIXME: Return the value */
+                       /* FIXME: Return the value in ENDFILTER */
                        MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
-                       /* Return the saved arp_pfs value to call_filter */
-                       ia64_mov (code, IA64_R9, cfg->arch.reg_saved_ar_pfs);
+
+                       /* Load the return address */
                        ia64_adds_imm (code, GP_SCRATCH_REG, spvar->inst_offset, cfg->frame_reg);
                        ia64_ld8_hint (code, GP_SCRATCH_REG, GP_SCRATCH_REG, 0);
+
+                       /* Test caller */
+                       ia64_cmp_eq (code, 6, 7, GP_SCRATCH_REG, IA64_R0);
+                       ia64_br_cond_pred (code, 7, 4);
+
+                       /* Called by call_filter */
+                       /* Pop frame */
+                       ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
+                       ia64_mov_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
+                       ia64_br_ret_reg (code, IA64_B0);                        
+
+                       /* Called by CALL_HANDLER */
                        ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
                        ia64_br_cond_reg (code, IA64_B6);
                        break;
@@ -3230,51 +3370,14 @@ static guint64 nops_for_ins_types [6] = {
 #endif
 
 static void
-ia64_emit_bundle_manual (Ia64CodegenState *code)
-{
-       int j, ins_type, template;
-
-       if (code->nins == 0)
-               return;
-
-       g_assert (code->nins == 3 || ((code->nins == 2) && code->itypes [1] == IA64_INS_TYPE_LX));
-
-       /* Verify template is correct */
-       template = code->template;
-       for (j = 0; j < 3; ++j) {
-               if (code->stops [j])
-                       g_assert (stops_in_template [template]);
-
-               ins_type = ins_types_in_template [template][j];
-               switch (code->itypes [j]) {
-               case IA64_INS_TYPE_A:
-                       g_assert ((ins_type == IA64_INS_TYPE_I) || (ins_type == IA64_INS_TYPE_M));
-                       break;
-               case IA64_INS_TYPE_LX:
-                       g_assert (j == 1);
-                       g_assert (ins_type == IA64_INS_TYPE_LX);
-                       j ++;
-                       break;
-               default:
-                       g_assert (ins_type == code->itypes [j]);
-               }
-       }
-
-       ia64_emit_bundle_template (code, template, code->instructions [0], code->instructions [1], code->instructions [2]);
-       code->template = 0;
-       code->nins = 0;
-       code->dep_info_pos = 0;
-}
-
-static void
-ia64_analyze_deps (Ia64CodegenState *code, int *deps_start)
+ia64_analyze_deps (Ia64CodegenState *code, int *deps_start, int *stops)
 {
        int i, pos, ins_index, current_deps_start, current_ins_start, reg;
        guint8 *deps = code->dep_info;
        gboolean need_stop, no_stop;
 
        for (i = 0; i < code->nins; ++i)
-               code->stops [i] = FALSE;
+               stops [i] = FALSE;
        
        ins_index = 0;
        current_deps_start = 0;
@@ -3413,7 +3516,7 @@ ia64_analyze_deps (Ia64CodegenState *code, int *deps_start)
 
                if (need_stop && !no_stop) {
                        g_assert (ins_index > 0);
-                       code->stops [ins_index - 1] = 1;
+                       stops [ins_index - 1] = 1;
 
                        DEBUG_INS_SCHED (printf ("STOP\n"));
                        current_deps_start = current_ins_start;
@@ -3426,18 +3529,19 @@ ia64_analyze_deps (Ia64CodegenState *code, int *deps_start)
 
        if (code->nins > 0) {
                /* No dependency info for the last instruction */
-               code->stops [code->nins - 1] = 1;
+               stops [code->nins - 1] = 1;
        }
 
        deps_start [code->nins] = code->dep_info_pos;
 }
 
 static void
-ia64_real_emit_bundle (Ia64CodegenState *code, int *deps_start, int n, guint64 template, guint64 ins1, guint64 ins2, guint64 ins3, guint8 nops)
+ia64_real_emit_bundle (Ia64CodegenState *code, int *deps_start, int *stops, int n, guint64 template, guint64 ins1, guint64 ins2, guint64 ins3, guint8 nops)
 {
-       g_assert (n <= code->nins);
        int stop_pos, i, deps_to_shift, dep_shift;
 
+       g_assert (n <= code->nins);
+
        // if (n > 1) printf ("FOUND: %ld.\n", template);
 
        ia64_emit_bundle_template (code, template, ins1, ins2, ins3);
@@ -3472,6 +3576,30 @@ ia64_real_emit_bundle (Ia64CodegenState *code, int *deps_start, int n, guint64 t
        for (i = 0; i < code->nins + 1 - n; ++i)
                deps_start [i] = deps_start [n + i] - dep_shift;
 
+       /* Determine the exact positions of instructions with unwind ops */
+       if (code->unw_op_count) {
+               int ins_pos [16];
+               int curr_ins, curr_ins_pos;
+
+               curr_ins = 0;
+               curr_ins_pos = ((code->buf - code->region_start - 16) / 16) * 3;
+               for (i = 0; i < 3; ++i) {
+                       if (! (nops & (1 << i))) {
+                               ins_pos [curr_ins] = curr_ins_pos + i;
+                               curr_ins ++;
+                       }
+               }
+
+               for (i = code->unw_op_pos; i < code->unw_op_count; ++i) {
+                       if (code->unw_ops_pos [i] < n) {
+                               code->unw_ops [i].when = ins_pos [code->unw_ops_pos [i]];
+                               //printf ("UNW-OP: %d -> %d\n", code->unw_ops_pos [i], code->unw_ops [i].when);
+                       }
+               }
+               if (code->unw_op_pos < code->unw_op_count)
+                       code->unw_op_pos += n;
+       }
+
        if (n == code->nins) {
                code->template = 0;
                code->nins = 0;
@@ -3479,7 +3607,7 @@ ia64_real_emit_bundle (Ia64CodegenState *code, int *deps_start, int n, guint64 t
        else {
                memcpy (&code->instructions [0], &code->instructions [n], (code->nins - n) * sizeof (guint64));
                memcpy (&code->itypes [0], &code->itypes [n], (code->nins - n) * sizeof (int));
-               memcpy (&code->stops [0], &code->stops [n], (code->nins - n) * sizeof (int));
+               memcpy (&stops [0], &stops [n], (code->nins - n) * sizeof (int));
                code->nins -= n;
        }
 }
@@ -3489,26 +3617,21 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
 {
        int i, ins_type, template, nins_to_emit;
        int deps_start [16];
+       int stops [16];
        gboolean found;
 
-       if (!code->automatic) {
-               ia64_emit_bundle_manual (code);
-               return;
-       }
-
        /*
         * We implement a simple scheduler which tries to put three instructions 
         * per bundle, then two, then one.
         */
-
-       ia64_analyze_deps (code, deps_start);
+       ia64_analyze_deps (code, deps_start, stops);
 
        if ((code->nins >= 3) && !code->one_ins_per_bundle) {
                /* Find a suitable template */
                for (template = 0; template < 32; ++template) {
-                       if (stops_in_template [template][0] != code->stops [0] ||
-                               stops_in_template [template][1] != code->stops [1] ||
-                               stops_in_template [template][2] != code->stops [2])
+                       if (stops_in_template [template][0] != stops [0] ||
+                               stops_in_template [template][1] != stops [1] ||
+                               stops_in_template [template][2] != stops [2])
                                continue;
 
                        found = TRUE;
@@ -3528,7 +3651,7 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
                                found = debug_ins_sched ();
 
                        if (found) {
-                               ia64_real_emit_bundle (code, deps_start, 3, template, code->instructions [0], code->instructions [1], code->instructions [2], 0);
+                               ia64_real_emit_bundle (code, deps_start, stops, 3, template, code->instructions [0], code->instructions [1], code->instructions [2], 0);
                                break;
                        }
                }
@@ -3542,9 +3665,9 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
        if ((code->nins >= 2) && !code->one_ins_per_bundle) {
                /* Try a nop at the end */
                for (template = 0; template < 32; ++template) {
-                       if (stops_in_template [template][0] != code->stops [0] ||
-                               ((stops_in_template [template][1] != code->stops [1]) &&
-                                (stops_in_template [template][2] != code->stops [1])))
+                       if (stops_in_template [template][0] != stops [0] ||
+                               ((stops_in_template [template][1] != stops [1]) &&
+                                (stops_in_template [template][2] != stops [1])))
                                 
                                continue;
 
@@ -3555,7 +3678,7 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
                        if (!debug_ins_sched ())
                                continue;
 
-                       ia64_real_emit_bundle (code, deps_start, 2, template, code->instructions [0], code->instructions [1], nops_for_ins_types [ins_types_in_template [template][2]], 1 << 2);
+                       ia64_real_emit_bundle (code, deps_start, stops, 2, template, code->instructions [0], code->instructions [1], nops_for_ins_types [ins_types_in_template [template][2]], 1 << 2);
                        break;
                }
        }
@@ -3567,9 +3690,9 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
        if ((code->nins >= 2) && !code->one_ins_per_bundle) {
                /* Try a nop in the middle */
                for (template = 0; template < 32; ++template) {
-                       if (((stops_in_template [template][0] != code->stops [0]) &&
-                                (stops_in_template [template][1] != code->stops [0])) ||
-                               stops_in_template [template][2] != code->stops [1])
+                       if (((stops_in_template [template][0] != stops [0]) &&
+                                (stops_in_template [template][1] != stops [0])) ||
+                               stops_in_template [template][2] != stops [1])
                                continue;
 
                        if (!ITYPE_MATCH (ins_types_in_template [template][0], code->itypes [0]) ||
@@ -3579,7 +3702,7 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
                        if (!debug_ins_sched ())
                                continue;
 
-                       ia64_real_emit_bundle (code, deps_start, 2, template, code->instructions [0], nops_for_ins_types [ins_types_in_template [template][1]], code->instructions [1], 1 << 1);
+                       ia64_real_emit_bundle (code, deps_start, stops, 2, template, code->instructions [0], nops_for_ins_types [ins_types_in_template [template][1]], code->instructions [1], 1 << 1);
                        break;
                }
        }
@@ -3587,8 +3710,8 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
        if ((code->nins >= 2) && flush && !code->one_ins_per_bundle) {
                /* Try a nop at the beginning */
                for (template = 0; template < 32; ++template) {
-                       if ((stops_in_template [template][1] != code->stops [0]) ||
-                               (stops_in_template [template][2] != code->stops [1]))
+                       if ((stops_in_template [template][1] != stops [0]) ||
+                               (stops_in_template [template][2] != stops [1]))
                                continue;
 
                        if (!ITYPE_MATCH (ins_types_in_template [template][1], code->itypes [0]) ||
@@ -3598,7 +3721,7 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
                        if (!debug_ins_sched ())
                                continue;
 
-                       ia64_real_emit_bundle (code, deps_start, 2, template, nops_for_ins_types [ins_types_in_template [template][0]], code->instructions [0], code->instructions [1], 1 << 0);
+                       ia64_real_emit_bundle (code, deps_start, stops, 2, template, nops_for_ins_types [ins_types_in_template [template][0]], code->instructions [0], code->instructions [1], 1 << 0);
                        break;
                }
        }
@@ -3614,43 +3737,43 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
 
        while (nins_to_emit > 0) {
                if (!debug_ins_sched ())
-                       code->stops [0] = 1;
+                       stops [0] = 1;
                switch (code->itypes [0]) {
                case IA64_INS_TYPE_A:
-                       if (code->stops [0])
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
+                       if (stops [0])
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
                        else
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
                        break;
                case IA64_INS_TYPE_I:
-                       if (code->stops [0])
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MIIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
+                       if (stops [0])
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
                        else
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MII, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
                        break;
                case IA64_INS_TYPE_M:
-                       if (code->stops [0])
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
+                       if (stops [0])
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
                        else
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
                        break;
                case IA64_INS_TYPE_B:
-                       if (code->stops [0])
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MIBS, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
+                       if (stops [0])
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIBS, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
                        else
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MIB, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIB, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
                        break;
                case IA64_INS_TYPE_F:
-                       if (code->stops [0])
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MFIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
+                       if (stops [0])
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
                        else
-                               ia64_real_emit_bundle (code, deps_start, 1, IA64_TEMPLATE_MFI, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
+                               ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFI, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
                        break;
                case IA64_INS_TYPE_LX:
-                       if (code->stops [0] || code->stops [1])
-                               ia64_real_emit_bundle (code, deps_start, 2, IA64_TEMPLATE_MLXS, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
+                       if (stops [0] || stops [1])
+                               ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLXS, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
                        else
-                               ia64_real_emit_bundle (code, deps_start, 2, IA64_TEMPLATE_MLX, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
+                               ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLX, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
                        nins_to_emit --;
                        break;
                default:
@@ -3660,6 +3783,22 @@ ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
        }
 }
 
+unw_dyn_region_info_t*
+mono_ia64_create_unwind_region (Ia64CodegenState *code)
+{
+       unw_dyn_region_info_t *r;
+
+       g_assert (code->nins == 0);
+       r = g_malloc0 (_U_dyn_region_info_size (code->unw_op_count));
+       memcpy (&r->op, &code->unw_ops, sizeof (unw_dyn_op_t) * code->unw_op_count);
+       r->op_count = code->unw_op_count;
+       r->insn_count = ((code->buf - code->region_start) >> 4) * 3;
+       code->unw_op_count = 0;
+       code->unw_op_pos = 0;
+       code->region_start = code->buf;
+
+       return r;
+}
 
 static void 
 ia64_patch (unsigned char* code, gpointer target)
@@ -3782,6 +3921,8 @@ mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, Mono
 
                target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
 
+               if (patch_info->type == MONO_PATCH_INFO_NONE)
+                       continue;
                if (mono_compile_aot) {
                        NOT_IMPLEMENTED;
                }
@@ -3799,8 +3940,6 @@ mono_arch_emit_prolog (MonoCompile *cfg)
        int alloc_size, pos, i;
        Ia64CodegenState code;
        CallInfo *cinfo;
-       unw_dyn_region_info_t *r_pro;
-       int unw_op_count;
        
        sig = mono_method_signature (method);
        pos = 0;
@@ -3817,7 +3956,6 @@ mono_arch_emit_prolog (MonoCompile *cfg)
        cfg->native_code = g_malloc (cfg->code_size);
 
        ia64_codegen_init (code, cfg->native_code);
-       ia64_codegen_set_automatic (code, FALSE);
 
        alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
        if (cfg->param_area)
@@ -3841,63 +3979,53 @@ mono_arch_emit_prolog (MonoCompile *cfg)
 
        alloc_size -= pos;
 
-       /* Initialize unwind info */
-       r_pro = g_malloc0 (_U_dyn_region_info_size (3));
-       unw_op_count = 0;
-
-       ia64_begin_bundle_template (code, IA64_TEMPLATE_MIIS);
+       ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + cfg->arch.reg_saved_ar_pfs);
        ia64_alloc (code, cfg->arch.reg_saved_ar_pfs, cfg->arch.reg_local0 - cfg->arch.reg_in0, cfg->arch.reg_out0 - cfg->arch.reg_local0, cfg->arch.n_out_regs, 0);
+       ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + cfg->arch.reg_saved_b0);
        ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
 
-       _U_dyn_op_save_reg (&r_pro->op[unw_op_count++], _U_QP_TRUE, /* when=*/ 0,
-                                               /* reg=*/ UNW_IA64_AR_PFS, /* dst=*/ UNW_IA64_GR + cfg->arch.reg_saved_ar_pfs);
-       _U_dyn_op_save_reg (&r_pro->op[unw_op_count++], _U_QP_TRUE, /* when=*/ 1,
-                                               /* reg=*/ UNW_IA64_RP, /* dst=*/ UNW_IA64_GR + cfg->arch.reg_saved_b0);
-
        if ((alloc_size || cinfo->stack_usage) && !cfg->arch.omit_fp) {
-               ia64_mov (code, cfg->frame_reg, IA64_SP);
-               _U_dyn_op_save_reg (&r_pro->op[unw_op_count++], _U_QP_TRUE, /* when=*/ 2,
-                                                       /* reg=*/ UNW_IA64_SP, /* dst=*/ UNW_IA64_GR + cfg->frame_reg);
-       }
-       else {
-               if (cfg->arch.omit_fp && alloc_size && ia64_is_imm14 (-alloc_size)) {
-                       /* FIXME: Add unwind info */
-                       ia64_adds_imm (code, IA64_SP, (-alloc_size), IA64_SP);
-               }
-               else
-                       ia64_nop_i (code, 0);
+               ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + cfg->arch.reg_saved_sp);
+               ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
+               if (cfg->frame_reg != cfg->arch.reg_saved_sp)
+                       ia64_mov (code, cfg->frame_reg, IA64_SP);
        }
-       ia64_stop (code);
-       ia64_begin_bundle (code);
-
-       /* Finish unwind info */
-       r_pro->op_count = unw_op_count;
-       r_pro->insn_count = (code.buf - cfg->native_code) >> 4;
-
-       cfg->arch.r_pro = r_pro;
-
-       ia64_codegen_set_automatic (code, TRUE);
 
        if (alloc_size) {
-               /* See mono_emit_stack_alloc */
-#if defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
-               NOT_IMPLEMENTED;
-#else
+               int pagesize = getpagesize ();
 
+#if defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
+               if (alloc_size >= pagesize) {
+                       gint32 remaining_size = alloc_size;
+
+                       /* Generate stack touching code */
+                       ia64_mov (code, GP_SCRATCH_REG, IA64_SP);                       
+                       while (remaining_size >= pagesize) {
+                               ia64_movl (code, GP_SCRATCH_REG2, pagesize);
+                               ia64_sub (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
+                               ia64_ld8 (code, GP_SCRATCH_REG2, GP_SCRATCH_REG);
+                               remaining_size -= pagesize;
+                       }
+               }
+#endif
                if (ia64_is_imm14 (-alloc_size)) {
                        if (cfg->arch.omit_fp)
-                               /* Already done */
-                               ;
-                       else
-                               ia64_adds_imm (code, IA64_SP, (-alloc_size), IA64_SP);
+                               ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
+                       ia64_adds_imm (code, IA64_SP, (-alloc_size), IA64_SP);
                }
                else {
-                       ia64_movl (code, GP_SCRATCH_REG, -alloc_size); ia64_stop (code);
+                       ia64_movl (code, GP_SCRATCH_REG, -alloc_size);
+                       if (cfg->arch.omit_fp)
+                               ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
                        ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
                }
-#endif
        }
 
+       ia64_begin_bundle (code);
+
+       /* Initialize unwind info */
+       cfg->arch.r_pro = mono_ia64_create_unwind_region (&code);
+
        if (sig->ret->type != MONO_TYPE_VOID) {
                if ((cinfo->ret.storage == ArgInIReg) && (cfg->ret->opcode != OP_REGVAR)) {
                        /* Save volatile arguments to the stack */
@@ -4081,43 +4209,37 @@ mono_arch_emit_epilog (MonoCompile *cfg)
        g_free (cinfo);
 
        ia64_begin_bundle (code);
-       ia64_codegen_set_automatic (code, FALSE);
-
-       if (cfg->arch.stack_alloc_size && cfg->arch.omit_fp && !ia64_is_imm14 (cfg->arch.stack_alloc_size)) {
-               ia64_begin_bundle_template (code, IA64_TEMPLATE_MLXS);
-               ia64_nop_m (code, 0);
-               ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
-               ia64_stop (code);
-               ia64_begin_bundle (code);
-       }
 
-       ia64_begin_bundle_template (code, IA64_TEMPLATE_MII);
+       code.region_start = cfg->native_code;
+
+       /* Label the unwind state at the start of the exception throwing region */
+       //ia64_unw_label_state (code, 1234);
+
        if (cfg->arch.stack_alloc_size) {
                if (cfg->arch.omit_fp) {
-                       if (ia64_is_imm14 (cfg->arch.stack_alloc_size))
+                       if (ia64_is_imm14 (cfg->arch.stack_alloc_size)) {
+                               ia64_unw_pop_frames (code, 1);
                                ia64_adds_imm (code, IA64_SP, (cfg->arch.stack_alloc_size), IA64_SP);
-                       else
+                       } else {
+                               ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
+                               ia64_unw_pop_frames (code, 1);
                                ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
+                       }
                }
-               else
+               else {
+                       ia64_unw_pop_frames (code, 1);
                        ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
+               }
        }
-       else
-               /* FIXME: Optimize this away */
-               ia64_nop_m (code, 0);
        ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
        ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
-       ia64_begin_bundle (code);
-
-       ia64_begin_bundle_template (code, IA64_TEMPLATE_BBBS);
        ia64_br_ret_reg (code, IA64_B0);
-       ia64_nop_b (code, 0);
-       ia64_nop_b (code, 0); ia64_stop (code);
-       ia64_begin_bundle (code);
 
-       ia64_codegen_set_automatic (code, TRUE);
        ia64_codegen_close (code);
 
+       cfg->arch.r_epilog = mono_ia64_create_unwind_region (&code);
+       cfg->arch.r_pro->next = cfg->arch.r_epilog;
+
        cfg->code_len = code.buf - cfg->native_code;
 
        g_assert (cfg->code_len < cfg->code_size);
@@ -4127,13 +4249,12 @@ void
 mono_arch_emit_exceptions (MonoCompile *cfg)
 {
        MonoJumpInfo *patch_info;
-       int nthrows;
+       int i, nthrows;
        Ia64CodegenState code;
        gboolean empty = TRUE;
-       /*
+       //unw_dyn_region_info_t *r_exceptions;
        MonoClass *exc_classes [16];
        guint8 *exc_throw_start [16], *exc_throw_end [16];
-       */
        guint32 code_size = 0;
 
        /* Compute needed space */
@@ -4146,6 +4267,9 @@ mono_arch_emit_exceptions (MonoCompile *cfg)
                        code_size += 4 + 7; /* sizeof (float) + alignment */
        }
 
+       if (code_size == 0)
+               return;
+
        while (cfg->code_len + code_size > (cfg->code_size - 16)) {
                cfg->code_size *= 2;
                cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
@@ -4154,6 +4278,9 @@ mono_arch_emit_exceptions (MonoCompile *cfg)
 
        ia64_codegen_init (code, cfg->native_code + cfg->code_len);
 
+       /* The unwind state here is the same as before the epilog */
+       //ia64_unw_copy_state (code, 1234);
+
        /* add code to raise exceptions */
        /* FIXME: Optimize this */
        nthrows = 0;
@@ -4163,34 +4290,77 @@ mono_arch_emit_exceptions (MonoCompile *cfg)
                        MonoClass *exc_class;
                        guint8* throw_ip;
                        guint8* buf;
+                       guint64 exc_token_index;
 
                        exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
                        g_assert (exc_class);
+                       exc_token_index = mono_metadata_token_index (exc_class->type_token);
                        throw_ip = cfg->native_code + patch_info->ip.i;
 
                        ia64_begin_bundle (code);
 
                        ia64_patch (cfg->native_code + patch_info->ip.i, code.buf);
 
-                       ia64_movl (code, cfg->arch.reg_out0 + 0, exc_class->type_token);
+                       /* Find a throw sequence for the same exception class */
+                       for (i = 0; i < nthrows; ++i)
+                               if (exc_classes [i] == exc_class)
+                                       break;
 
-                       ia64_begin_bundle (code);
+                       if (i < nthrows) {
+                               gint64 offset = exc_throw_end [i] - 16 - throw_ip;
 
-                       patch_info->data.name = "mono_arch_throw_corlib_exception";
-                       patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
-                       patch_info->ip.i = code.buf - cfg->native_code;
+                               if (ia64_is_adds_imm (offset))
+                                       ia64_adds_imm (code, cfg->arch.reg_out0 + 1, offset, IA64_R0);
+                               else
+                                       ia64_movl (code, cfg->arch.reg_out0 + 1, offset);
 
-                       /* Indirect call */
-                       ia64_movl (code, GP_SCRATCH_REG, 0);
-                       ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
-                       ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
-                       ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
+                               buf = code.buf + code.nins;
+                               ia64_br_cond_pred (code, 0, 0);
+                               ia64_begin_bundle (code);
+                               ia64_patch (buf, exc_throw_start [i]);
+
+                               patch_info->type = MONO_PATCH_INFO_NONE;
+                       }
+                       else {
+                               /* Arg1 */
+                               buf = code.buf;
+                               ia64_movl (code, cfg->arch.reg_out0 + 1, 0);
 
-                       /* Compute the offset */
-                       buf = code.buf + 32;
-                       ia64_movl (code, cfg->arch.reg_out0 + 1, buf - throw_ip);
+                               ia64_begin_bundle (code);
 
-                       ia64_br_call_reg (code, IA64_B0, IA64_B6);
+                               if (nthrows < 16) {
+                                       exc_classes [nthrows] = exc_class;
+                                       exc_throw_start [nthrows] = code.buf;
+                               }
+
+                               /* Arg2 */
+                               if (ia64_is_adds_imm (exc_token_index))
+                                       ia64_adds_imm (code, cfg->arch.reg_out0 + 0, exc_token_index, IA64_R0);
+                               else
+                                       ia64_movl (code, cfg->arch.reg_out0 + 0, exc_token_index);
+
+                               patch_info->data.name = "mono_arch_throw_corlib_exception";
+                               patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
+                               patch_info->ip.i = code.buf + code.nins - cfg->native_code;
+
+                               /* Indirect call */
+                               ia64_movl (code, GP_SCRATCH_REG, 0);
+                               ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
+                               ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
+                               ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
+
+                               ia64_br_call_reg (code, IA64_B0, IA64_B6);
+
+                               /* Patch up the throw offset */
+                               ia64_begin_bundle (code);
+
+                               ia64_patch (buf, (gpointer)(code.buf - 16 - throw_ip));
+
+                               if (nthrows < 16) {
+                                       exc_throw_end [nthrows] = code.buf;
+                                       nthrows ++;
+                               }
+                       }
 
                        empty = FALSE;
                        break;
@@ -4206,6 +4376,10 @@ mono_arch_emit_exceptions (MonoCompile *cfg)
 
        ia64_codegen_close (code);
 
+       /* FIXME: */
+       //r_exceptions = mono_ia64_create_unwind_region (&code);
+       //cfg->arch.r_epilog = r_exceptions;
+
        cfg->code_len = code.buf - cfg->native_code;
 
        g_assert (cfg->code_len < cfg->code_size);
@@ -4376,6 +4550,18 @@ mono_arch_save_unwind_info (MonoCompile *cfg)
        di->u.pi.regions = cfg->arch.r_pro;
 
        _U_dyn_register (di);
+
+       /*
+       {
+               unw_dyn_region_info_t *region = di->u.pi.regions;
+
+               printf ("Unwind info for method %s:\n", mono_method_full_name (cfg->method, TRUE));
+               while (region) {
+                       printf ("    [Region: %d]\n", region->insn_count);
+                       region = region->next;
+               }
+       }
+       */
 }
 
 void
@@ -4508,16 +4694,6 @@ mono_arch_get_delegate_method_ptr_addr (guint8* code, gpointer *regs)
 
 static gboolean tls_offset_inited = FALSE;
 
-#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
-
-static void
-setup_stack (MonoJitTlsData *tls)
-{
-       NOT_IMPLEMENTED;
-}
-
-#endif
-
 void
 mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
 {
@@ -4527,26 +4703,11 @@ mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
                appdomain_tls_offset = mono_domain_get_tls_offset ();
                thread_tls_offset = mono_thread_get_tls_offset ();
        }               
-
-#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
-       setup_stack (tls);
-#endif
 }
 
 void
 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
 {
-#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
-       struct sigaltstack sa;
-
-       sa.ss_sp = tls->signal_stack;
-       sa.ss_size = SIGNAL_STACK_SIZE;
-       sa.ss_flags = SS_DISABLE;
-       sigaltstack  (&sa, NULL);
-
-       if (tls->signal_stack)
-               munmap (tls->signal_stack, SIGNAL_STACK_SIZE);
-#endif
 }
 
 void
@@ -4576,7 +4737,7 @@ mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_re
                        vtarg->dreg = mono_regstate_next_int (cfg->rs);
                        mono_bblock_add_inst (cfg->cbb, vtarg);
 
-                       mono_call_inst_add_outarg_reg (call, vtarg->dreg, out_reg, FALSE);
+                       mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, out_reg, FALSE);
 
                        out_reg ++;
                }
@@ -4593,7 +4754,7 @@ mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_re
                this->dreg = mono_regstate_next_int (cfg->rs);
                mono_bblock_add_inst (cfg->cbb, this);
 
-               mono_call_inst_add_outarg_reg (call, this->dreg, out_reg, FALSE);
+               mono_call_inst_add_outarg_reg (cfg, call, this->dreg, out_reg, FALSE);
        }
 }
 
@@ -4602,89 +4763,37 @@ mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethod
 {
        MonoInst *ins = NULL;
 
-       /* FIXME: */
-       return NULL;
-
-       if (cmethod->klass == mono_defaults.math_class) {
-               if (strcmp (cmethod->name, "Sin") == 0) {
-                       MONO_INST_NEW (cfg, ins, OP_SIN);
-                       ins->inst_i0 = args [0];
-               } else if (strcmp (cmethod->name, "Cos") == 0) {
-                       MONO_INST_NEW (cfg, ins, OP_COS);
-                       ins->inst_i0 = args [0];
-               } else if (strcmp (cmethod->name, "Tan") == 0) {
-                               return ins;
-                       MONO_INST_NEW (cfg, ins, OP_TAN);
-                       ins->inst_i0 = args [0];
-               } else if (strcmp (cmethod->name, "Atan") == 0) {
-                               return ins;
-                       MONO_INST_NEW (cfg, ins, OP_ATAN);
-                       ins->inst_i0 = args [0];
-               } else if (strcmp (cmethod->name, "Sqrt") == 0) {
-                       MONO_INST_NEW (cfg, ins, OP_SQRT);
-                       ins->inst_i0 = args [0];
-               } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
-                       MONO_INST_NEW (cfg, ins, OP_ABS);
-                       ins->inst_i0 = args [0];
-               }
-#if 0
-               /* OP_FREM is not IEEE compatible */
-               else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
-                       MONO_INST_NEW (cfg, ins, OP_FREM);
-                       ins->inst_i0 = args [0];
-                       ins->inst_i1 = args [1];
-               }
-#endif
+       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) {
-                       MonoInst *ins_iconst;
                        guint32 opcode;
 
                        if (fsig->params [0]->type == MONO_TYPE_I4)
-                               opcode = OP_ATOMIC_ADD_NEW_I4;
+                               opcode = OP_ATOMIC_ADD_IMM_NEW_I4;
                        else if (fsig->params [0]->type == MONO_TYPE_I8)
-                               opcode = OP_ATOMIC_ADD_NEW_I8;
+                               opcode = OP_ATOMIC_ADD_IMM_NEW_I8;
                        else
                                g_assert_not_reached ();
                        MONO_INST_NEW (cfg, ins, opcode);
-                       MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
-                       ins_iconst->inst_c0 = 1;
-
+                       ins->inst_imm = 1;
                        ins->inst_i0 = args [0];
-                       ins->inst_i1 = ins_iconst;
                } else if (strcmp (cmethod->name, "Decrement") == 0) {
-                       MonoInst *ins_iconst;
                        guint32 opcode;
 
                        if (fsig->params [0]->type == MONO_TYPE_I4)
-                               opcode = OP_ATOMIC_ADD_NEW_I4;
+                               opcode = OP_ATOMIC_ADD_IMM_NEW_I4;
                        else if (fsig->params [0]->type == MONO_TYPE_I8)
-                               opcode = OP_ATOMIC_ADD_NEW_I8;
+                               opcode = OP_ATOMIC_ADD_IMM_NEW_I8;
                        else
                                g_assert_not_reached ();
                        MONO_INST_NEW (cfg, ins, opcode);
-                       MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
-                       ins_iconst->inst_c0 = -1;
-
+                       ins->inst_imm = -1;
                        ins->inst_i0 = args [0];
-                       ins->inst_i1 = ins_iconst;
-               } else if (strcmp (cmethod->name, "Add") == 0) {
-                       guint32 opcode;
-
-                       if (fsig->params [0]->type == MONO_TYPE_I4)
-                               opcode = OP_ATOMIC_ADD_I4;
-                       else if (fsig->params [0]->type == MONO_TYPE_I8)
-                               opcode = OP_ATOMIC_ADD_I8;
-                       else
-                               g_assert_not_reached ();
-                       
-                       MONO_INST_NEW (cfg, ins, opcode);
-
-                       ins->inst_i0 = args [0];
-                       ins->inst_i1 = args [1];
                } else if (strcmp (cmethod->name, "Exchange") == 0) {
                        guint32 opcode;
 
@@ -4699,6 +4808,20 @@ mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethod
 
                        MONO_INST_NEW (cfg, ins, opcode);
 
+                       ins->inst_i0 = args [0];
+                       ins->inst_i1 = args [1];
+               } else if (strcmp (cmethod->name, "Add") == 0) {
+                       guint32 opcode;
+
+                       if (fsig->params [0]->type == MONO_TYPE_I4)
+                               opcode = OP_ATOMIC_ADD_NEW_I4;
+                       else if (fsig->params [0]->type == MONO_TYPE_I8)
+                               opcode = OP_ATOMIC_ADD_NEW_I8;
+                       else
+                               g_assert_not_reached ();
+                       
+                       MONO_INST_NEW (cfg, ins, opcode);
+
                        ins->inst_i0 = args [0];
                        ins->inst_i1 = args [1];
                } else if (strcmp (cmethod->name, "Read") == 0 && (fsig->params [0]->type == MONO_TYPE_I8)) {
@@ -4706,11 +4829,6 @@ mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethod
                        MONO_INST_NEW (cfg, ins, CEE_LDIND_I8);
                        ins->inst_i0 = args [0];
                }
-
-               /* 
-                * Can't implement CompareExchange methods this way since they have
-                * three arguments.
-                */
        }
 
        return ins;