2009-02-18 Rodrigo Kumpera <rkumpera@novell.com>
[mono.git] / mono / mini / mini-sparc.c
index 1552e50c16ecd0ce4aaded00570504bcfff37b74..e44f8ffdb579bc89458cdd7e86df65b082c06225 100644 (file)
 #include <thread.h>
 #endif
 
+#include <unistd.h>
+#include <sys/mman.h>
+
 #include <mono/metadata/appdomain.h>
 #include <mono/metadata/debug-helpers.h>
+#include <mono/metadata/tokentype.h>
 #include <mono/utils/mono-math.h>
 
 #include "mini-sparc.h"
-#include "inssel.h"
 #include "trace.h"
 #include "cpu-sparc.h"
+#include "jit-icalls.h"
+#include "ir-emit.h"
 
 /*
  * Sparc V9 means two things:
  * Register usage:
  * - %i0..%i<n> hold the incoming arguments, these are never written by JITted 
  * code. Unused input registers are used for global register allocation.
- * - %l0..%l7 is used for local register allocation
- * - %o0..%o6 is used for outgoing arguments
+ * - %o0..%o5 and %l7 is used for local register allocation and passing arguments
+ * - %l0..%l6 is used for global register allocation
  * - %o7 and %g1 is used as scratch registers in opcodes
  * - all floating point registers are used for local register allocation except %f0. 
  *   Only double precision registers are used.
+ * In 64 bit mode:
+ * - fp registers %d0..%d30 are used for parameter passing, and %d32..%d62 are
+ *   used for local allocation.
  */
 
 /*
  *      This test casts (guint64)-1 to double and then back to guint64 again.
  *    Under x86, it returns 0, while under sparc it returns -1.
  *
- * In addition to this, the runtime requires the truncl function, or its 
+ * In addition to this, the runtime requires the trunc function, or its 
  * solaris counterpart, aintl, to do some double->int conversions. If this 
  * function is not available, it is emulated somewhat, but the results can be
  * strange.
  */
 
+/*
+ * SPARCV9 FIXME:
+ * - optimize sparc_set according to the memory model
+ * - when non-AOT compiling, compute patch targets immediately so we don't
+ *   have to emit the 6 byte template.
+ * - varags
+ * - struct arguments/returns
+ */
+
+/*
+ * SPARCV9 ISSUES:
+ * - sparc_call_simple can't be used in a lot of places since the displacement
+ *   might not fit into an imm30.
+ * - g1 can't be used in a lot of places since it is used as a scratch reg in
+ *   sparc_set.
+ * - sparc_f0 can't be used as a scratch register on V9
+ * - the %d34..%d62 fp registers are encoded as: %dx = %f(x - 32 + 1), ie.
+ *   %d36 = %f5.
+ * - ldind.i4/u4 needs to sign extend/clear out upper word -> slows things down
+ * - ins->dreg can't be used as a scatch register in r4 opcodes since it might
+ *   be a double precision register which has no single precision part.
+ * - passing/returning structs is hard to implement, because:
+ *   - the spec is very hard to understand
+ *   - it requires knowledge about the fields of structure, needs to handle
+ *     nested structures etc.
+ */
+
 /*
  * Possible optimizations:
  * - delay slot scheduling
  * - allocate large constants to registers
- * - use %o registers for local allocation
- * - implement unwinding through native frames
  * - add more mul/div/rem optimizations
  */
 
-#if SPARCV9
-#error "Sparc V9 support not yet implemented."
-#endif
-
 #ifndef __linux__
 #define MONO_SPARC_THR_TLS 1
 #endif
 
-#define NOT_IMPLEMENTED g_assert_not_reached ();
+/*
+ * There was a 64 bit bug in glib-2.2: g_bit_nth_msf (0, -1) would return 32,
+ * causing infinite loops in dominator computation. So glib-2.4 is required.
+ */
+#ifdef SPARCV9
+#if GLIB_MAJOR_VERSION == 2 && GLIB_MINOR_VERSION < 4
+#error "glib 2.4 or later is required for 64 bit mode."
+#endif
+#endif
 
 #define ALIGN_TO(val,align) (((val) + ((align) - 1)) & ~((align) - 1))
 
 #define SIGNAL_STACK_SIZE (64 * 1024)
 
+#define STACK_BIAS MONO_SPARC_STACK_BIAS
+
+#ifdef SPARCV9
+
+/* %g1 is used by sparc_set */
+#define GP_SCRATCH_REG sparc_g4
+/* %f0 is used for parameter passing */
+#define FP_SCRATCH_REG sparc_f30
+#define ARGS_OFFSET (STACK_BIAS + 128)
+
+#else
+
+#define FP_SCRATCH_REG sparc_f0
+#define ARGS_OFFSET 68
+#define GP_SCRATCH_REG sparc_g1
+
+#endif
+
 /* Whenever the CPU supports v9 instructions */
-gboolean sparcv9 = FALSE;
+static gboolean sparcv9 = FALSE;
 
-static gpointer mono_arch_get_lmf_addr (void);
+/* Whenever this is a 64bit executable */
+#if SPARCV9
+static gboolean v64 = TRUE;
+#else
+static gboolean v64 = FALSE;
+#endif
 
-static int
-mono_spillvar_offset_float (MonoCompile *cfg, int spillvar);
+static gpointer mono_arch_get_lmf_addr (void);
 
 const char*
 mono_arch_regname (int reg) {
@@ -119,11 +178,48 @@ mono_arch_regname (int reg) {
        return "unknown";
 }
 
+const char*
+mono_arch_fregname (int reg) {
+       static const char *rnames [] = {
+               "sparc_f0", "sparc_f1", "sparc_f2", "sparc_f3", "sparc_f4", 
+               "sparc_f5", "sparc_f6", "sparc_f7", "sparc_f8", "sparc_f9",
+               "sparc_f10", "sparc_f11", "sparc_f12", "sparc_f13", "sparc_f14", 
+               "sparc_f15", "sparc_f16", "sparc_f17", "sparc_f18", "sparc_f19",
+               "sparc_f20", "sparc_f21", "sparc_f22", "sparc_f23", "sparc_f24", 
+               "sparc_f25", "sparc_f26", "sparc_f27", "sparc_f28", "sparc_f29",
+               "sparc_f30", "sparc_f31"
+       };
+
+       if (reg >= 0 && reg < 32)
+               return rnames [reg];
+       else
+               return "unknown";
+}
+
 /*
  * Initialize the cpu to execute managed code.
  */
 void
 mono_arch_cpu_init (void)
+{
+       guint32 dummy;
+       /* make sure sparcv9 is initialized for embedded use */
+       mono_arch_cpu_optimizazions(&dummy);
+}
+
+/*
+ * Initialize architecture specific code.
+ */
+void
+mono_arch_init (void)
+{
+}
+
+/*
+ * Cleanup architecture specific code.
+ */
+void
+mono_arch_cleanup (void)
 {
 }
 
@@ -168,11 +264,6 @@ mono_arch_cpu_optimizazions (guint32 *exclude_mask)
        return opts;
 }
 
-static void
-mono_sparc_break (void)
-{
-}
-
 #ifdef __GNUC__
 #define flushi(addr)    __asm__ __volatile__ ("iflush %0"::"r"(addr):"memory")
 #else /* assume Sun's compiler */
@@ -182,6 +273,10 @@ static void flushi(void *addr)
 }
 #endif
 
+#ifndef __linux__
+void sync_instruction_memory(caddr_t addr, int len);
+#endif
+
 void
 mono_arch_flush_icache (guint8 *code, gint size)
 {
@@ -189,18 +284,28 @@ mono_arch_flush_icache (guint8 *code, gint size)
        /* Hopefully this is optimized based on the actual CPU */
        sync_instruction_memory (code, size);
 #else
-       guint64 *p = (guint64*)code;
-       guint64 *end = (guint64*)(code + ((size + 8) /8));
-
-       /* 
-        * FIXME: Flushing code in dword chunks in _slow_.
+       gulong start = (gulong) code;
+       gulong end = start + size;
+       gulong align;
+
+       /* Sparcv9 chips only need flushes on 32 byte
+        * cacheline boundaries.
+        *
+        * Sparcv8 needs a flush every 8 bytes.
         */
-       while (p < end)
+       align = (sparcv9 ? 32 : 8);
+
+       start &= ~(align - 1);
+       end = (end + (align - 1)) & ~(align - 1);
+
+       while (start < end) {
 #ifdef __GNUC__
-               __asm__ __volatile__ ("iflush %0"::"r"(p++));
+               __asm__ __volatile__ ("iflush %0"::"r"(start));
 #else
-                       flushi (p ++);
+               flushi (start);
 #endif
+               start += align;
+       }
 #endif
 }
 
@@ -228,6 +333,8 @@ mono_sparc_flushw (void)
 
                g_assert ((code - start) < 64);
 
+               mono_arch_flush_icache ((guint8*)start, (guint8*)code - (guint8*)start);
+
                flushw = (gpointer)start;
 
                inited = 1;
@@ -242,11 +349,22 @@ mono_arch_flush_register_windows (void)
        mono_sparc_flushw ();
 }
 
+gboolean 
+mono_arch_is_inst_imm (gint64 imm)
+{
+       return sparc_is_imm13 (imm);
+}
+
 gboolean 
 mono_sparc_is_v9 (void) {
        return sparcv9;
 }
 
+gboolean 
+mono_sparc_is_sparc64 (void) {
+       return v64;
+}
+
 typedef enum {
        ArgInIReg,
        ArgInIRegPair,
@@ -254,7 +372,9 @@ typedef enum {
        ArgInFReg,
        ArgInFRegPair,
        ArgOnStack,
-       ArgOnStackPair
+       ArgOnStackPair,
+       ArgInFloatReg,  /* V9 only */
+       ArgInDoubleReg  /* V9 only */
 } ArgStorage;
 
 typedef struct {
@@ -295,7 +415,7 @@ add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean pair)
                }
 
                /* Allways reserve stack space for parameters passed in registers */
-               (*stack_size) += 4;
+               (*stack_size) += sizeof (gpointer);
        }
        else {
                if (*gr < PARAM_REGS - 1) {
@@ -307,50 +427,106 @@ add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean pair)
                else if (*gr >= PARAM_REGS) {
                        /* A pair of stack locations */
                        ainfo->storage = ArgOnStackPair;
-                       ainfo->offset = *stack_size;
                }
                else {
                        ainfo->storage = ArgInSplitRegStack;
                        ainfo->reg = *gr;
-                       ainfo->offset = *stack_size;
                        (*gr) ++;
                }
 
-               (*stack_size) += 8;
+               (*stack_size) += 2 * sizeof (gpointer);
+       }
+}
+
+#ifdef SPARCV9
+
+#define FLOAT_PARAM_REGS 32
+
+static void inline
+add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean single)
+{
+       ainfo->offset = *stack_size;
+
+       if (single) {
+               if (*gr >= FLOAT_PARAM_REGS) {
+                       ainfo->storage = ArgOnStack;
+               }
+               else {
+                       /* A single is passed in an even numbered fp register */
+                       ainfo->storage = ArgInFloatReg;
+                       ainfo->reg = *gr + 1;
+                       (*gr) += 2;
+               }
+       }
+       else {
+               if (*gr < FLOAT_PARAM_REGS) {
+                       /* A double register */
+                       ainfo->storage = ArgInDoubleReg;
+                       ainfo->reg = *gr;
+                       (*gr) += 2;
+               }
+               else {
+                       ainfo->storage = ArgOnStack;
+               }
        }
+
+       (*stack_size) += sizeof (gpointer);
 }
 
+#endif
+
 /*
  * get_call_info:
  *
  *  Obtain information about a call according to the calling convention.
- * See the "System V ABI, Sparc Processor Supplement" Sparc V8 version document for
- * more information.
+ * For V8, see the "System V ABI, Sparc Processor Supplement" Sparc V8 version 
+ * document for more information.
+ * For V9, see the "Low Level System Information (64-bit psABI)" chapter in
+ * the 'Sparc Compliance Definition 2.4' document.
  */
 static CallInfo*
-get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
+get_call_info (MonoCompile *cfg, MonoMethodSignature *sig, gboolean is_pinvoke)
 {
-       guint32 i, gr, simpletype;
+       guint32 i, gr, fr;
        int n = sig->hasthis + sig->param_count;
        guint32 stack_size = 0;
        CallInfo *cinfo;
+       MonoType *ret_type;
+       MonoGenericSharingContext *gsctx = cfg ? cfg->generic_sharing_context : NULL;
 
        cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
 
        gr = 0;
+       fr = 0;
+
+#ifdef SPARCV9
+       if (MONO_TYPE_ISSTRUCT ((sig->ret))) {
+               /* The address of the return value is passed in %o0 */
+               add_general (&gr, &stack_size, &cinfo->ret, FALSE);
+               cinfo->ret.reg += sparc_i0;
+       }
+#endif
 
        /* this */
        if (sig->hasthis)
                add_general (&gr, &stack_size, cinfo->args + 0, FALSE);
 
+       if ((sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
+               gr = PARAM_REGS;
+
+               /* Emit the signature cookie just before the implicit arguments */
+               add_general (&gr, &stack_size, &cinfo->sig_cookie, FALSE);
+       }
+
        for (i = 0; i < sig->param_count; ++i) {
                ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
+               MonoType *ptype;
 
                if ((sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
+                       gr = PARAM_REGS;
+
                        /* Emit the signature cookie just before the implicit arguments */
                        add_general (&gr, &stack_size, &cinfo->sig_cookie, FALSE);
-                       /* Prevent implicit arguments from being passed in registers */
-                       gr = PARAM_REGS;
                }
 
                DEBUG(printf("param %d: ", i));
@@ -360,28 +536,33 @@ get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
                        add_general (&gr, &stack_size, ainfo, FALSE);
                        continue;
                }
-               simpletype = sig->params [i]->type;
-       enum_calc_size:
-               switch (simpletype) {
+               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:
                case MONO_TYPE_U1:
                        add_general (&gr, &stack_size, ainfo, FALSE);
                        /* the value is in the ls byte */
-                       ainfo->offset += 3;
+                       ainfo->offset += sizeof (gpointer) - 1;
                        break;
                case MONO_TYPE_I2:
                case MONO_TYPE_U2:
                case MONO_TYPE_CHAR:
                        add_general (&gr, &stack_size, ainfo, FALSE);
                        /* the value is in the ls word */
-                       ainfo->offset += 2;
+                       ainfo->offset += sizeof (gpointer) - 2;
                        break;
                case MONO_TYPE_I4:
                case MONO_TYPE_U4:
+                       add_general (&gr, &stack_size, ainfo, FALSE);
+                       /* the value is in the ls dword */
+                       ainfo->offset += sizeof (gpointer) - 4;
+                       break;
                case MONO_TYPE_I:
                case MONO_TYPE_U:
                case MONO_TYPE_PTR:
+               case MONO_TYPE_FNPTR:
                case MONO_TYPE_CLASS:
                case MONO_TYPE_OBJECT:
                case MONO_TYPE_STRING:
@@ -389,12 +570,17 @@ get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
                case MONO_TYPE_ARRAY:
                        add_general (&gr, &stack_size, ainfo, FALSE);
                        break;
-               case MONO_TYPE_VALUETYPE:
-                       if (sig->params [i]->data.klass->enumtype) {
-                               simpletype = sig->params [i]->data.klass->enum_basetype->type;
-                               goto enum_calc_size;
+               case MONO_TYPE_GENERICINST:
+                       if (!mono_type_generic_inst_is_valuetype (sig->params [i])) {
+                               add_general (&gr, &stack_size, ainfo, FALSE);
+                               break;
                        }
-
+                       /* Fall through */
+               case MONO_TYPE_VALUETYPE:
+#ifdef SPARCV9
+                       if (sig->pinvoke)
+                               NOT_IMPLEMENTED;
+#endif
                        add_general (&gr, &stack_size, ainfo, FALSE);
                        break;
                case MONO_TYPE_TYPEDBYREF:
@@ -402,109 +588,128 @@ get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
                        break;
                case MONO_TYPE_U8:
                case MONO_TYPE_I8:
+#ifdef SPARCV9
+                       add_general (&gr, &stack_size, ainfo, FALSE);
+#else
                        add_general (&gr, &stack_size, ainfo, TRUE);
+#endif
                        break;
                case MONO_TYPE_R4:
+#ifdef SPARCV9
+                       add_float (&fr, &stack_size, ainfo, TRUE);
+                       gr ++;
+#else
                        /* single precision values are passed in integer registers */
                        add_general (&gr, &stack_size, ainfo, FALSE);
+#endif
                        break;
                case MONO_TYPE_R8:
+#ifdef SPARCV9
+                       add_float (&fr, &stack_size, ainfo, FALSE);
+                       gr ++;
+#else
                        /* double precision values are passed in a pair of registers */
                        add_general (&gr, &stack_size, ainfo, TRUE);
+#endif
                        break;
                default:
                        g_assert_not_reached ();
                }
        }
 
-       /* return value */
-       {
-               simpletype = sig->ret->type;
-enum_retvalue:
-               switch (simpletype) {
-               case MONO_TYPE_BOOLEAN:
-               case MONO_TYPE_I1:
-               case MONO_TYPE_U1:
-               case MONO_TYPE_I2:
-               case MONO_TYPE_U2:
-               case MONO_TYPE_CHAR:
-               case MONO_TYPE_I4:
-               case MONO_TYPE_U4:
-               case MONO_TYPE_I:
-               case MONO_TYPE_U:
-               case MONO_TYPE_PTR:
-               case MONO_TYPE_CLASS:
-               case MONO_TYPE_OBJECT:
-               case MONO_TYPE_SZARRAY:
-               case MONO_TYPE_ARRAY:
-               case MONO_TYPE_STRING:
-                       cinfo->ret.storage = ArgInIReg;
-                       cinfo->ret.reg = sparc_i0;
-                       if (gr < 1)
-                               gr = 1;
-                       break;
-               case MONO_TYPE_U8:
-               case MONO_TYPE_I8:
-                       cinfo->ret.storage = ArgInIRegPair;
-                       cinfo->ret.reg = sparc_i0;
-                       if (gr < 2)
-                               gr = 2;
-                       break;
-               case MONO_TYPE_R4:
-               case MONO_TYPE_R8:
-                       cinfo->ret.storage = ArgInFReg;
-                       cinfo->ret.reg = sparc_f0;
-                       break;
-               case MONO_TYPE_VALUETYPE:
-                       if (sig->ret->data.klass->enumtype) {
-                               simpletype = sig->ret->data.klass->enum_basetype->type;
-                               goto enum_retvalue;
-                       }
-                       cinfo->ret.storage = ArgOnStack;
-                       break;
-               case MONO_TYPE_TYPEDBYREF:
-                       cinfo->ret.storage = ArgOnStack;
-                       break;
-               case MONO_TYPE_VOID:
-                       break;
-               default:
-                       g_error ("Can't handle as return value 0x%x", sig->ret->type);
-               }
-       }
+       if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
+               gr = PARAM_REGS;
 
-       cinfo->stack_usage = stack_size;
-       cinfo->reg_usage = gr;
-       return cinfo;
-}
+               /* Emit the signature cookie just before the implicit arguments */
+               add_general (&gr, &stack_size, &cinfo->sig_cookie, FALSE);
+       }
 
-static gboolean
-is_regsize_var (MonoType *t) {
-       if (t->byref)
-               return TRUE;
-       switch (t->type) {
+       /* 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_CHAR:
        case MONO_TYPE_I1:
        case MONO_TYPE_U1:
        case MONO_TYPE_I2:
        case MONO_TYPE_U2:
+       case MONO_TYPE_CHAR:
        case MONO_TYPE_I4:
        case MONO_TYPE_U4:
        case MONO_TYPE_I:
        case MONO_TYPE_U:
-               return TRUE;
-       case MONO_TYPE_OBJECT:
-       case MONO_TYPE_STRING:
+       case MONO_TYPE_PTR:
+       case MONO_TYPE_FNPTR:
        case MONO_TYPE_CLASS:
+       case MONO_TYPE_OBJECT:
        case MONO_TYPE_SZARRAY:
        case MONO_TYPE_ARRAY:
-               return TRUE;
+       case MONO_TYPE_STRING:
+               cinfo->ret.storage = ArgInIReg;
+               cinfo->ret.reg = sparc_i0;
+               if (gr < 1)
+                       gr = 1;
+               break;
+       case MONO_TYPE_U8:
+       case MONO_TYPE_I8:
+#ifdef SPARCV9
+               cinfo->ret.storage = ArgInIReg;
+               cinfo->ret.reg = sparc_i0;
+               if (gr < 1)
+                       gr = 1;
+#else
+               cinfo->ret.storage = ArgInIRegPair;
+               cinfo->ret.reg = sparc_i0;
+               if (gr < 2)
+                       gr = 2;
+#endif
+               break;
+       case MONO_TYPE_R4:
+       case MONO_TYPE_R8:
+               cinfo->ret.storage = ArgInFReg;
+               cinfo->ret.reg = sparc_f0;
+               break;
+       case MONO_TYPE_GENERICINST:
+               if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
+                       cinfo->ret.storage = ArgInIReg;
+                       cinfo->ret.reg = sparc_i0;
+                       if (gr < 1)
+                               gr = 1;
+                       break;
+               }
+               /* Fall through */
        case MONO_TYPE_VALUETYPE:
-               if (t->data.klass->enumtype)
-                       return is_regsize_var (t->data.klass->enum_basetype);
-               return FALSE;
+               if (v64) {
+                       if (sig->pinvoke)
+                               NOT_IMPLEMENTED;
+                       else
+                               /* Already done */
+                               ;
+               }
+               else
+                       cinfo->ret.storage = ArgOnStack;
+               break;
+       case MONO_TYPE_TYPEDBYREF:
+               if (v64) {
+                       if (sig->pinvoke)
+                               /* Same as a valuetype with size 24 */
+                               NOT_IMPLEMENTED;
+                       else
+                               /* Already done */
+                               ;
+               }
+               else
+                       cinfo->ret.storage = ArgOnStack;
+               break;
+       case MONO_TYPE_VOID:
+               break;
+       default:
+               g_error ("Can't handle as return value 0x%x", sig->ret->type);
        }
-       return FALSE;
+
+       cinfo->stack_usage = stack_size;
+       cinfo->reg_usage = gr;
+       return cinfo;
 }
 
 GList *
@@ -530,8 +735,7 @@ mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
                if (ins->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT) || (ins->opcode == OP_REGVAR) || (ins->opcode == OP_ARG))
                        continue;
 
-               /* we can only allocate 32 bit values */
-               if (is_regsize_var (ins->inst_vtype)) {
+               if (mono_is_regsize_var (ins->inst_vtype)) {
                        g_assert (MONO_VARINFO (cfg, i)->reg == -1);
                        g_assert (i == vmv->idx);
 
@@ -550,16 +754,16 @@ mono_arch_get_global_int_regs (MonoCompile *cfg)
        MonoMethodSignature *sig;
        CallInfo *cinfo;
 
-       sig = cfg->method->signature;
+       sig = mono_method_signature (cfg->method);
 
-       cinfo = get_call_info (sig, FALSE);
+       cinfo = get_call_info (cfg, sig, FALSE);
 
        /* Use unused input registers */
        for (i = cinfo->reg_usage; i < 6; ++i)
                regs = g_list_prepend (regs, GUINT_TO_POINTER (sparc_i0 + i));
 
-       /* Use %l0..%l3 as global registers */
-       for (i = 16; i < 20; ++i)
+       /* Use %l0..%l6 as global registers */
+       for (i = sparc_l0; i < sparc_l7; ++i)
                regs = g_list_prepend (regs, GUINT_TO_POINTER (i));
 
        g_free (cinfo);
@@ -584,8 +788,9 @@ mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
  * Set var information according to the calling convention. sparc version.
  * The locals var stuff should most likely be split in another method.
  */
+
 void
-mono_arch_allocate_vars (MonoCompile *m)
+mono_arch_allocate_vars (MonoCompile *cfg)
 {
        MonoMethodSignature *sig;
        MonoMethodHeader *header;
@@ -593,38 +798,56 @@ mono_arch_allocate_vars (MonoCompile *m)
        int i, offset, size, align, curinst;
        CallInfo *cinfo;
 
-       header = ((MonoMethodNormal *)m->method)->header;
+       header = mono_method_get_header (cfg->method);
 
-       sig = m->method->signature;
+       sig = mono_method_signature (cfg->method);
 
-       cinfo = get_call_info (sig, FALSE);
+       cinfo = get_call_info (cfg, sig, FALSE);
 
        if (sig->ret->type != MONO_TYPE_VOID) {
                switch (cinfo->ret.storage) {
                case ArgInIReg:
                case ArgInFReg:
+                       cfg->ret->opcode = OP_REGVAR;
+                       cfg->ret->inst_c0 = cinfo->ret.reg;
+                       break;
                case ArgInIRegPair:
-                       m->ret->opcode = OP_REGVAR;
-                       m->ret->inst_c0 = cinfo->ret.reg;
+                       if (((sig->ret->type == MONO_TYPE_I8) || (sig->ret->type == MONO_TYPE_U8))) {
+                               MonoInst *low = get_vreg_to_inst (cfg, cfg->ret->dreg + 1);
+                               MonoInst *high = get_vreg_to_inst (cfg, cfg->ret->dreg + 2);
+
+                               low->opcode = OP_REGVAR;
+                               low->dreg = cinfo->ret.reg + 1;
+                               high->opcode = OP_REGVAR;
+                               high->dreg = cinfo->ret.reg;
+                       }
+                       cfg->ret->opcode = OP_REGVAR;
+                       cfg->ret->inst_c0 = cinfo->ret.reg;
                        break;
                case ArgOnStack:
+#ifdef SPARCV9
+                       g_assert_not_reached ();
+#else
                        /* valuetypes */
-                       m->ret->opcode = OP_REGOFFSET;
-                       m->ret->inst_basereg = sparc_fp;
-                       m->ret->inst_offset = 64;
+                       cfg->vret_addr->opcode = OP_REGOFFSET;
+                       cfg->vret_addr->inst_basereg = sparc_fp;
+                       cfg->vret_addr->inst_offset = 64;
+#endif
                        break;
                default:
                        NOT_IMPLEMENTED;
                }
+               cfg->ret->dreg = cfg->ret->inst_c0;
        }
 
        /*
-        * We use the Sparc V8 calling conventions for managed code as well.
+        * We use the ABI calling conventions for managed code as well.
+        * Exception: valuetypes are never returned in registers on V9.
         * FIXME: Use something more optimized.
         */
 
        /* Locals are allocated backwards from %fp */
-       m->frame_reg = sparc_fp;
+       cfg->frame_reg = sparc_fp;
        offset = 0;
 
        /* 
@@ -632,28 +855,31 @@ mono_arch_allocate_vars (MonoCompile *m)
         * handling.
         */
        if (header->num_clauses)
-               offset += 8;
+               offset += sizeof (gpointer) * 2;
 
-       if (m->method->save_lmf) {
+       if (cfg->method->save_lmf) {
                offset += sizeof (MonoLMF);
-               m->arch.lmf_offset = offset;
+               cfg->arch.lmf_offset = offset;
        }
 
-       curinst = m->locals_start;
-       for (i = curinst; i < m->num_varinfo; ++i) {
-               inst = m->varinfo [i];
+       curinst = cfg->locals_start;
+       for (i = curinst; i < cfg->num_varinfo; ++i) {
+               inst = cfg->varinfo [i];
 
-               if (inst->opcode == OP_REGVAR) {
+               if ((inst->opcode == OP_REGVAR) || (inst->opcode == OP_REGOFFSET)) {
                        //g_print ("allocating local %d to %s\n", i, mono_arch_regname (inst->dreg));
                        continue;
                }
 
-               /* inst->unused indicates native sized value types, this is used by the
+               if (inst->flags & MONO_INST_IS_DEAD)
+                       continue;
+
+               /* inst->backend.is_pinvoke indicates native sized value types, this is used by the
                * pinvoke wrappers when they call functions returning structure */
-               if (inst->unused && MONO_TYPE_ISSTRUCT (inst->inst_vtype) && inst->inst_vtype->type != MONO_TYPE_TYPEDBYREF)
+               if (inst->backend.is_pinvoke && MONO_TYPE_ISSTRUCT (inst->inst_vtype) && inst->inst_vtype->type != MONO_TYPE_TYPEDBYREF)
                        size = mono_class_native_size (inst->inst_vtype->data.klass, &align);
                else
-                       size = mono_type_stack_size (inst->inst_vtype, &align);
+                       size = mini_type_stack_size (cfg->generic_sharing_context, inst->inst_vtype, &align);
 
                /* 
                 * This is needed since structures containing doubles must be doubleword 
@@ -673,27 +899,29 @@ mono_arch_allocate_vars (MonoCompile *m)
                offset &= ~(align - 1);
                inst->opcode = OP_REGOFFSET;
                inst->inst_basereg = sparc_fp;
-               inst->inst_offset = -offset;
+               inst->inst_offset = STACK_BIAS + -offset;
 
                //g_print ("allocating local %d to [%s - %d]\n", i, mono_arch_regname (inst->inst_basereg), - inst->inst_offset);
        }
 
        if (sig->call_convention == MONO_CALL_VARARG) {
-               m->sig_cookie = cinfo->sig_cookie.offset + 68;
+               cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
        }
 
        for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
-               inst = m->varinfo [i];
+               inst = cfg->args [i];
                if (inst->opcode != OP_REGVAR) {
                        ArgInfo *ainfo = &cinfo->args [i];
                        gboolean inreg = TRUE;
                        MonoType *arg_type;
+                       ArgStorage storage;
 
                        if (sig->hasthis && (i == 0))
                                arg_type = &mono_defaults.object_class->byval_arg;
                        else
                                arg_type = sig->params [i - sig->hasthis];
 
+#ifndef SPARCV9
                        if (!arg_type->byref && ((arg_type->type == MONO_TYPE_R4) 
                                                                         || (arg_type->type == MONO_TYPE_R8)))
                                /*
@@ -701,8 +929,10 @@ mono_arch_allocate_vars (MonoCompile *m)
                                 * save them to the stack in the prolog.
                                 */
                                inreg = FALSE;
+#endif
 
                        /* FIXME: Allocate volatile arguments to registers */
+                       /* FIXME: This makes the argument holding a vtype address into volatile */
                        if (inst->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
                                inreg = FALSE;
 
@@ -710,17 +940,38 @@ mono_arch_allocate_vars (MonoCompile *m)
                                /* FIXME: this isn't needed */
                                inreg = FALSE;
 
-                       switch (ainfo->storage) {
+                       inst->opcode = OP_REGOFFSET;
+
+                       if (!inreg)
+                               storage = ArgOnStack;
+                       else
+                               storage = ainfo->storage;
+
+                       switch (storage) {
                        case ArgInIReg:
+                               inst->opcode = OP_REGVAR;
+                               inst->dreg = sparc_i0 + ainfo->reg;
+                               break;
                        case ArgInIRegPair:
-                               if (inreg) {
-                                       inst->opcode = OP_REGVAR;
-                                       inst->dreg = sparc_i0 + ainfo->reg;
-                                       break;
-                               }
-                               else {
-                                       /* Fall through */
+                               if (inst->type == STACK_I8) {
+                                       MonoInst *low = get_vreg_to_inst (cfg, inst->dreg + 1);
+                                       MonoInst *high = get_vreg_to_inst (cfg, inst->dreg + 2);
+
+                                       low->opcode = OP_REGVAR;
+                                       low->dreg = sparc_i0 + ainfo->reg + 1;
+                                       high->opcode = OP_REGVAR;
+                                       high->dreg = sparc_i0 + ainfo->reg;
                                }
+                               inst->opcode = OP_REGVAR;
+                               inst->dreg = sparc_i0 + ainfo->reg;                                     
+                               break;
+                       case ArgInFloatReg:
+                       case ArgInDoubleReg:
+                               /* 
+                                * Since float regs are volatile, we save the arguments to
+                                * the stack in the prolog.
+                                * FIXME: Avoid this if the method contains no calls.
+                                */
                        case ArgOnStack:
                        case ArgOnStackPair:
                        case ArgInSplitRegStack:
@@ -728,7 +979,7 @@ mono_arch_allocate_vars (MonoCompile *m)
                                inst->opcode = OP_REGOFFSET;
                                /* in parent frame */
                                inst->inst_basereg = sparc_fp;
-                               inst->inst_offset = ainfo->offset + 68;
+                               inst->inst_offset = ainfo->offset + ARGS_OFFSET;
 
                                if (!arg_type->byref && (arg_type->type == MONO_TYPE_R8)) {
                                        /* 
@@ -736,13 +987,13 @@ mono_arch_allocate_vars (MonoCompile *m)
                                         * memory locations. So if the offset is misaligned, we copy the
                                         * argument to a stack location in the prolog.
                                         */
-                                       if (inst->inst_offset % 8) {
+                                       if ((inst->inst_offset - STACK_BIAS) % 8) {
                                                inst->inst_basereg = sparc_fp;
                                                offset += 8;
                                                align = 8;
                                                offset += align - 1;
                                                offset &= ~(align - 1);
-                                               inst->inst_offset = -offset;
+                                               inst->inst_offset = STACK_BIAS + -offset;
 
                                        }
                                }
@@ -758,253 +1009,400 @@ mono_arch_allocate_vars (MonoCompile *m)
                                 * are destructively modified in a lot of places in inssel.brg.
                                 */
                                MonoInst *indir;
-                               MONO_INST_NEW (m, indir, 0);
+                               MONO_INST_NEW (cfg, indir, 0);
                                *indir = *inst;
-                               inst->opcode = OP_SPARC_INARG_VT;
+                               inst->opcode = OP_VTARG_ADDR;
                                inst->inst_left = indir;
                        }
                }
        }
 
+       /* Add a properly aligned dword for use by int<->float conversion opcodes */
+       offset += 8;
+       offset = ALIGN_TO (offset, 8);
+       cfg->arch.float_spill_slot_offset = offset;
+
        /* 
         * spillvars are stored between the normal locals and the storage reserved
         * by the ABI.
         */
 
-       m->stack_offset = offset;
-
-       /* Add a properly aligned dword for use by int<->float conversion opcodes */
-       m->spill_count ++;
-       mono_spillvar_offset_float (m, 0);
+       cfg->stack_offset = offset;
 
        g_free (cinfo);
 }
 
-/* 
- * 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.
- */
-MonoCallInst*
-mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual) {
-       MonoInst *arg, *in;
+void
+mono_arch_create_vars (MonoCompile *cfg)
+{
        MonoMethodSignature *sig;
-       int i, n;
-       CallInfo *cinfo;
-       ArgInfo *ainfo;
-       guint32 extra_space = 0;
 
-       sig = call->signature;
-       n = sig->param_count + sig->hasthis;
-       
-       cinfo = get_call_info (sig, sig->pinvoke);
+       sig = mono_method_signature (cfg->method);
 
-       for (i = 0; i < n; ++i) {
-               ainfo = cinfo->args + i;
-               if (is_virtual && i == 0) {
-                       /* the argument will be attached to the call instruction */
-                       in = call->args [i];
-               } else {
-                       if ((sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
-                               /* FIXME: Test varargs with 0 implicit args */
-                               /* FIXME: Test interaction with hasthis */
-                               /* Emit the signature cookie just before the first implicit argument */
-                               MonoInst *sig_arg;
-                               /* FIXME: Add support for signature tokens to AOT */
-                               cfg->disable_aot = TRUE;
-                               /* We allways pass the signature on the stack for simplicity */
-                               MONO_INST_NEW (cfg, arg, OP_SPARC_OUTARG_MEM);
-                               arg->inst_basereg = sparc_sp;
-                               arg->inst_imm = 68 + cinfo->sig_cookie.offset;
-                               MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
-                               sig_arg->inst_p0 = call->signature;
-                               arg->inst_left = sig_arg;
-                               arg->type = STACK_PTR;
-                               /* prepend, so they get reversed */
-                               arg->next = call->out_args;
-                               call->out_args = arg;
-                       }
+       if (MONO_TYPE_ISSTRUCT ((sig->ret))) {
+               cfg->vret_addr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_ARG);
+               if (G_UNLIKELY (cfg->verbose_level > 1)) {
+                       printf ("vret_addr = ");
+                       mono_print_ins (cfg->vret_addr);
+               }
+       }
+}
 
-                       MONO_INST_NEW (cfg, arg, OP_OUTARG);
-                       in = call->args [i];
-                       arg->cil_code = in->cil_code;
-                       arg->inst_left = in;
-                       arg->type = in->type;
-                       /* prepend, we'll need to reverse them later */
-                       arg->next = call->out_args;
-                       call->out_args = arg;
-
-                       if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(sig->params [i - sig->hasthis]))) {
-                               MonoInst *inst;
-                               gint align;
-                               guint32 offset, pad;
-                               guint32 size;
-                               
-                               if (sig->params [i - sig->hasthis]->type == MONO_TYPE_TYPEDBYREF) {
-                                       size = sizeof (MonoTypedRef);
-                                       align = 4;
-                               }
-                               else
-                               if (sig->pinvoke)
-                                       size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
-                               else
-                                       size = mono_type_stack_size (&in->klass->byval_arg, &align);
+static void
+add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, ArgStorage storage, int reg, guint32 sreg)
+{
+       MonoInst *arg;
 
-                               /* 
-                                * We use OP_OUTARG_VT to copy the valuetype to a stack location, then
-                                * use the normal OUTARG opcodes to pass the address of the location to
-                                * the callee.
-                                */
-                               MONO_INST_NEW (cfg, inst, OP_OUTARG_VT);
-                               inst->inst_left = in;
+       MONO_INST_NEW (cfg, arg, 0);
 
-                               /* The first 6 argument locations are reserved */
-                               if (cinfo->stack_usage < 24)
-                                       cinfo->stack_usage = 24;
+       arg->sreg1 = sreg;
 
-                               offset = ALIGN_TO (68 + cinfo->stack_usage, align);
-                               pad = offset - (68 + cinfo->stack_usage);
+       switch (storage) {
+       case ArgInIReg:
+               arg->opcode = OP_MOVE;
+               arg->dreg = mono_alloc_ireg (cfg);
 
-                               inst->inst_c1 = offset;
-                               inst->unused = size;
-                               arg->inst_left = inst;
+               mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, FALSE);
+               break;
+       case ArgInFloatReg:
+               arg->opcode = OP_FMOVE;
+               arg->dreg = mono_alloc_freg (cfg);
 
-                               cinfo->stack_usage += size;
-                               cinfo->stack_usage += pad;
-                       }
+               mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, TRUE);
+               break;
+       default:
+               g_assert_not_reached ();
+       }
 
-                       switch (ainfo->storage) {
-                       case ArgInIReg:
-                       case ArgInFReg:
-                       case ArgInIRegPair:
-                               if (ainfo->storage == ArgInIRegPair)
-                                       arg->opcode = OP_SPARC_OUTARG_REGPAIR;
-                               arg->unused = sparc_o0 + ainfo->reg;
-                               /* outgoing arguments begin at sp+68 */
-                               arg->inst_basereg = sparc_sp;
-                               arg->inst_imm = 68 + ainfo->offset;
-                               call->used_iregs |= 1 << ainfo->reg;
-
-                               if ((i >= sig->hasthis) && (sig->params [i - sig->hasthis]->type == MONO_TYPE_R8)) {
-                                       /*
-                                        * The OUTARG (freg) implementation needs an extra dword to store
-                                        * the temporary value.
-                                        */
-                                       extra_space += 8;
-                               }
-                               break;
-                       case ArgOnStack:
-                               arg->opcode = OP_SPARC_OUTARG_MEM;
-                               arg->inst_basereg = sparc_sp;
-                               arg->inst_imm = 68 + ainfo->offset;
-                               break;
-                       case ArgOnStackPair:
-                               arg->opcode = OP_SPARC_OUTARG_MEMPAIR;
-                               arg->inst_basereg = sparc_sp;
-                               arg->inst_imm = 68 + ainfo->offset;
-                               break;
-                       case ArgInSplitRegStack:
-                               arg->opcode = OP_SPARC_OUTARG_SPLIT_REG_STACK;
-                               arg->unused = sparc_o0 + ainfo->reg;
-                               arg->inst_basereg = sparc_sp;
-                               arg->inst_imm = 68 + ainfo->offset;
-                               call->used_iregs |= 1 << ainfo->reg;
-                               break;
-                       default:
-                               NOT_IMPLEMENTED;
-                       }
-               }
+       MONO_ADD_INS (cfg->cbb, arg);
+}
+
+static void
+add_outarg_load (MonoCompile *cfg, MonoCallInst *call, int opcode, int basereg, int offset, int reg)
+{
+       MonoInst *arg;
+       int dreg = mono_alloc_ireg (cfg);
+
+       EMIT_NEW_LOAD_MEMBASE (cfg, arg, OP_LOAD_MEMBASE, dreg, sparc_sp, offset);
+       MONO_ADD_INS (cfg->cbb, arg);
+
+       mono_call_inst_add_outarg_reg (cfg, call, dreg, reg, FALSE);
+}
+
+static void
+emit_pass_long (MonoCompile *cfg, MonoCallInst *call, ArgInfo *ainfo, MonoInst *in)
+{
+       int offset = ARGS_OFFSET + ainfo->offset;
+
+       switch (ainfo->storage) {
+       case ArgInIRegPair:
+               add_outarg_reg (cfg, call, ArgInIReg, sparc_o0 + ainfo->reg + 1, in->dreg + 1);
+               add_outarg_reg (cfg, call, ArgInIReg, sparc_o0 + ainfo->reg, in->dreg + 2);
+               break;
+       case ArgOnStackPair:
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI4_MEMBASE_REG, sparc_sp, offset, in->dreg + 2);
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI4_MEMBASE_REG, sparc_sp, offset + 4, in->dreg + 1);
+               break;
+       case ArgInSplitRegStack:
+               add_outarg_reg (cfg, call, ArgInIReg, sparc_o0 + ainfo->reg, in->dreg + 2);
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI4_MEMBASE_REG, sparc_sp, offset + 4, in->dreg + 1);
+               break;
+       default:
+               g_assert_not_reached ();
        }
+}
+
+static void
+emit_pass_double (MonoCompile *cfg, MonoCallInst *call, ArgInfo *ainfo, MonoInst *in)
+{
+       int offset = ARGS_OFFSET + ainfo->offset;
+
+       switch (ainfo->storage) {
+       case ArgInIRegPair:
+               /* floating-point <-> integer transfer must go through memory */
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, sparc_sp, offset, in->dreg);
+
+               /* Load into a register pair */
+               add_outarg_load (cfg, call, OP_LOADI4_MEMBASE, sparc_sp, offset, sparc_o0 + ainfo->reg);
+               add_outarg_load (cfg, call, OP_LOADI4_MEMBASE, sparc_sp, offset + 4, sparc_o0 + ainfo->reg + 1);
+               break;
+       case ArgOnStackPair:
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, sparc_sp, offset, in->dreg);
+               break;
+       case ArgInSplitRegStack:
+               /* floating-point <-> integer transfer must go through memory */
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, sparc_sp, offset, in->dreg);
+               /* Load most significant word into register */
+               add_outarg_load (cfg, call, OP_LOADI4_MEMBASE, sparc_sp, offset, sparc_o0 + ainfo->reg);
+               break;
+       default:
+               g_assert_not_reached ();
+       }
+}
+
+static void
+emit_pass_float (MonoCompile *cfg, MonoCallInst *call, ArgInfo *ainfo, MonoInst *in)
+{
+       int offset = ARGS_OFFSET + ainfo->offset;
+
+       switch (ainfo->storage) {
+       case ArgInIReg:
+               /* floating-point <-> integer transfer must go through memory */
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, sparc_sp, offset, in->dreg);
+               add_outarg_load (cfg, call, OP_LOADI4_MEMBASE, sparc_sp, offset, sparc_o0 + ainfo->reg);
+               break;
+       case ArgOnStack:
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, sparc_sp, offset, in->dreg);
+               break;
+       default:
+               g_assert_not_reached ();
+       }
+}
+
+static void
+emit_pass_other (MonoCompile *cfg, MonoCallInst *call, ArgInfo *ainfo, MonoType *arg_type, MonoInst *in);
+
+static void
+emit_pass_vtype (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo, ArgInfo *ainfo, MonoType *arg_type, MonoInst *in, gboolean pinvoke)
+{
+       MonoInst *arg;
+       guint32 align, offset, pad, size;
+
+       if (arg_type->type == MONO_TYPE_TYPEDBYREF) {
+               size = sizeof (MonoTypedRef);
+               align = sizeof (gpointer);
+       }
+       else if (pinvoke)
+               size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
+       else {
+               /* 
+                * Other backends use mono_type_stack_size (), but that
+                * aligns the size to 8, which is larger than the size of
+                * the source, leading to reads of invalid memory if the
+                * source is at the end of address space.
+                */
+               size = mono_class_value_size (in->klass, &align);
+       }
+
+       /* The first 6 argument locations are reserved */
+       if (cinfo->stack_usage < 6 * sizeof (gpointer))
+               cinfo->stack_usage = 6 * sizeof (gpointer);
+
+       offset = ALIGN_TO ((ARGS_OFFSET - STACK_BIAS) + cinfo->stack_usage, align);
+       pad = offset - ((ARGS_OFFSET - STACK_BIAS) + cinfo->stack_usage);
+
+       cinfo->stack_usage += size;
+       cinfo->stack_usage += pad;
+
+       /* 
+        * We use OP_OUTARG_VT to copy the valuetype to a stack location, then
+        * use the normal OUTARG opcodes to pass the address of the location to
+        * the callee.
+        */
+       if (size > 0) {
+               MONO_INST_NEW (cfg, arg, OP_OUTARG_VT);
+               arg->sreg1 = in->dreg;
+               arg->klass = in->klass;
+               arg->backend.size = size;
+               arg->inst_p0 = call;
+               arg->inst_p1 = mono_mempool_alloc (cfg->mempool, sizeof (ArgInfo));
+               memcpy (arg->inst_p1, ainfo, sizeof (ArgInfo));
+               ((ArgInfo*)(arg->inst_p1))->offset = STACK_BIAS + offset;
+               MONO_ADD_INS (cfg->cbb, arg);
+
+               MONO_INST_NEW (cfg, arg, OP_ADD_IMM);
+               arg->dreg = mono_alloc_preg (cfg);
+               arg->sreg1 = sparc_sp;
+               arg->inst_imm = STACK_BIAS + offset;
+               MONO_ADD_INS (cfg->cbb, arg);
+
+               emit_pass_other (cfg, call, ainfo, NULL, arg);
+       }
+}
+
+static void
+emit_pass_other (MonoCompile *cfg, MonoCallInst *call, ArgInfo *ainfo, MonoType *arg_type, MonoInst *in)
+{
+       int offset = ARGS_OFFSET + ainfo->offset;
+       int opcode;
+
+       switch (ainfo->storage) {
+       case ArgInIReg:
+               add_outarg_reg (cfg, call, ArgInIReg, sparc_o0 + ainfo->reg, in->dreg);
+               break;
+       case ArgOnStack:
+#ifdef SPARCV9
+               NOT_IMPLEMENTED;
+#else
+               if (offset & 0x1)
+                       opcode = OP_STOREI1_MEMBASE_REG;
+               else if (offset & 0x2)
+                       opcode = OP_STOREI2_MEMBASE_REG;
+               else
+                       opcode = OP_STOREI4_MEMBASE_REG;
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, opcode, sparc_sp, offset, in->dreg);
+#endif
+               break;
+       default:
+               g_assert_not_reached ();
+       }
+}
+
+static void
+emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
+{
+       MonoMethodSignature *tmp_sig;
 
        /*
-        * Reverse the call->out_args list.
+        * 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.
         */
-       {
-               MonoInst *prev = NULL, *list = call->out_args, *next;
-               while (list) {
-                       next = list->next;
-                       list->next = prev;
-                       prev = list;
-                       list = next;
+       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*));
+
+       /* FIXME: Add support for signature tokens to AOT */
+       cfg->disable_aot = TRUE;
+       /* We allways pass the signature on the stack for simplicity */
+       MONO_EMIT_NEW_STORE_MEMBASE_IMM (cfg, OP_STORE_MEMBASE_IMM, sparc_sp, ARGS_OFFSET + cinfo->sig_cookie.offset, tmp_sig);
+}
+
+void
+mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
+{
+       MonoInst *in;
+       MonoMethodSignature *sig;
+       int i, n;
+       CallInfo *cinfo;
+       ArgInfo *ainfo;
+       guint32 extra_space = 0;
+
+       sig = call->signature;
+       n = sig->param_count + sig->hasthis;
+       
+       cinfo = get_call_info (cfg, sig, sig->pinvoke);
+
+       if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
+               /* Set the 'struct/union return pointer' location on the stack */
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI4_MEMBASE_REG, sparc_sp, 64, call->vret_var->dreg);
+       }
+
+       for (i = 0; i < n; ++i) {
+               MonoType *arg_type;
+
+               ainfo = cinfo->args + i;
+
+               if ((sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
+                       /* Emit the signature cookie just before the first implicit argument */
+                       emit_sig_cookie (cfg, call, cinfo);
                }
-               call->out_args = prev;
+
+               in = call->args [i];
+
+               if (sig->hasthis && (i == 0))
+                       arg_type = &mono_defaults.object_class->byval_arg;
+               else
+                       arg_type = sig->params [i - sig->hasthis];
+
+               if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(sig->params [i - sig->hasthis])))
+                       emit_pass_vtype (cfg, call, cinfo, ainfo, arg_type, in, sig->pinvoke);
+               else if (!arg_type->byref && ((arg_type->type == MONO_TYPE_I8) || (arg_type->type == MONO_TYPE_U8)))
+                       emit_pass_long (cfg, call, ainfo, in);
+               else if (!arg_type->byref && (arg_type->type == MONO_TYPE_R8))
+                       emit_pass_double (cfg, call, ainfo, in);
+               else if (!arg_type->byref && (arg_type->type == MONO_TYPE_R4))
+                       emit_pass_float (cfg, call, ainfo, in);
+               else
+                       emit_pass_other (cfg, call, ainfo, arg_type, in);
        }
+
+       /* 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 + extra_space;
-       cfg->param_area = MAX (cfg->param_area, call->stack_usage);
-       cfg->flags |= MONO_CFG_HAS_CALLS;
 
        g_free (cinfo);
-       return call;
 }
 
+void
+mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
+{
+       ArgInfo *ainfo = (ArgInfo*)ins->inst_p1;
+       int size = ins->backend.size;
+
+       mini_emit_memcpy (cfg, sparc_sp, ainfo->offset, src->dreg, 0, size, 0);
+}
+
+void
+mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
+{
+       CallInfo *cinfo = get_call_info (cfg, mono_method_signature (method), FALSE);
+
+       switch (cinfo->ret.storage) {
+       case ArgInIReg:
+               MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg);
+               break;
+       case ArgInIRegPair:
+               MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg + 2);
+               MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg + 1, val->dreg + 1);
+               break;
+       case ArgInFReg:
+               if (mono_method_signature (method)->ret->type == MONO_TYPE_R4)
+                       MONO_EMIT_NEW_UNALU (cfg, OP_SETFRET, cfg->ret->dreg, val->dreg);
+               else
+                       MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
+               break;
+       default:
+               g_assert_not_reached ();
+       }
+
+       g_assert (cinfo);
+}
+
+int cond_to_sparc_cond [][3] = {
+       {sparc_be,   sparc_be,   sparc_fbe},
+       {sparc_bne,  sparc_bne,  0},
+       {sparc_ble,  sparc_ble,  sparc_fble},
+       {sparc_bge,  sparc_bge,  sparc_fbge},
+       {sparc_bl,   sparc_bl,   sparc_fbl},
+       {sparc_bg,   sparc_bg,   sparc_fbg},
+       {sparc_bleu, sparc_bleu, 0},
+       {sparc_beu,  sparc_beu,  0},
+       {sparc_blu,  sparc_blu,  sparc_fbl},
+       {sparc_bgu,  sparc_bgu,  sparc_fbg}
+};
+
 /* Map opcode to the sparc condition codes */
 static inline SparcCond
 opcode_to_sparc_cond (int opcode)
 {
+       CompRelation rel;
+       CompType t;
+
        switch (opcode) {
-       case OP_FBGE:
-               return sparc_fbge;
-       case OP_FBLE:
-               return sparc_fble;
-       case OP_FBEQ:
-       case OP_FCEQ:
-               return sparc_fbe;
-       case OP_FBLT:
-       case OP_FCLT:
-       case OP_FCLT_UN:
-               return sparc_fbl;
-       case OP_FBGT:
-       case OP_FCGT:
-       case OP_FCGT_UN:
-               return sparc_fbg;
-       case CEE_BEQ:
-       case OP_CEQ:
-       case OP_COND_EXC_EQ:
-               return sparc_be;
-       case CEE_BNE_UN:
-       case OP_COND_EXC_NE_UN:
-               return sparc_bne;
-       case CEE_BLT:
-       case OP_CLT:
-       case OP_COND_EXC_LT:
-               return sparc_bl;
-       case CEE_BLT_UN:
-       case OP_CLT_UN:
-       case OP_COND_EXC_LT_UN:
-               return sparc_blu;
-       case CEE_BGT:
-       case OP_CGT:
-       case OP_COND_EXC_GT:
-               return sparc_bg;
-       case CEE_BGT_UN:
-       case OP_CGT_UN:
-       case OP_COND_EXC_GT_UN:
-               return sparc_bgu;
-       case CEE_BGE:
-       case OP_COND_EXC_GE:
-               return sparc_bge;
-       case CEE_BGE_UN:
-       case OP_COND_EXC_GE_UN:
-               return sparc_beu;
-       case CEE_BLE:
-       case OP_COND_EXC_LE:
-               return sparc_ble;
-       case CEE_BLE_UN:
-       case OP_COND_EXC_LE_UN:
-               return sparc_bleu;
        case OP_COND_EXC_OV:
+       case OP_COND_EXC_IOV:
                return sparc_bvs;
        case OP_COND_EXC_C:
+       case OP_COND_EXC_IC:
                return sparc_bcs;
-
        case OP_COND_EXC_NO:
        case OP_COND_EXC_NC:
                NOT_IMPLEMENTED;
        default:
-               g_assert_not_reached ();
-               return sparc_be;
+               rel = mono_opcode_to_cond (opcode);
+               t = mono_opcode_to_type (opcode, -1);
+
+               return cond_to_sparc_cond [rel][t];
+               break;
        }
+
+       return -1;
 }
 
 #define COMPUTE_DISP(ins) \
@@ -1024,6 +1422,36 @@ if (ins->flags & MONO_INST_BRLABEL) { \
         } \
 }
 
+#ifdef SPARCV9
+#define DEFAULT_ICC sparc_xcc_short
+#else
+#define DEFAULT_ICC sparc_icc_short
+#endif
+
+#ifdef SPARCV9
+#define EMIT_COND_BRANCH_ICC(ins,cond,annul,filldelay,icc) \
+    do { \
+        gint32 disp; \
+        guint32 predict; \
+        COMPUTE_DISP(ins); \
+        predict = (disp != 0) ? 1 : 0; \
+        g_assert (sparc_is_imm19 (disp)); \
+        sparc_branchp (code, (annul), cond, icc, (predict), disp); \
+        if (filldelay) sparc_nop (code); \
+    } while (0)
+#define EMIT_COND_BRANCH(ins,cond,annul,filldelay) EMIT_COND_BRANCH_ICC ((ins), (cond), (annul), (filldelay), (sparc_xcc_short))
+#define EMIT_FLOAT_COND_BRANCH(ins,cond,annul,filldelay) \
+    do { \
+        gint32 disp; \
+        guint32 predict; \
+        COMPUTE_DISP(ins); \
+        predict = (disp != 0) ? 1 : 0; \
+        g_assert (sparc_is_imm19 (disp)); \
+        sparc_fbranch (code, (annul), cond, disp); \
+        if (filldelay) sparc_nop (code); \
+    } while (0)
+#else
+#define EMIT_COND_BRANCH_ICC(ins,cond,annul,filldelay,icc) g_assert_not_reached ()
 #define EMIT_COND_BRANCH_GENERAL(ins,bop,cond,annul,filldelay) \
     do { \
         gint32 disp; \
@@ -1032,9 +1460,9 @@ if (ins->flags & MONO_INST_BRLABEL) { \
         sparc_ ## bop (code, (annul), cond, disp); \
         if (filldelay) sparc_nop (code); \
     } while (0)
-
 #define EMIT_COND_BRANCH(ins,cond,annul,filldelay) EMIT_COND_BRANCH_GENERAL((ins),branch,(cond),annul,filldelay)
 #define EMIT_FLOAT_COND_BRANCH(ins,cond,annul,filldelay) EMIT_COND_BRANCH_GENERAL((ins),fbranch,(cond),annul,filldelay)
+#endif
 
 #define EMIT_COND_BRANCH_PREDICTED(ins,cond,annul,filldelay) \
     do { \
@@ -1043,7 +1471,7 @@ if (ins->flags & MONO_INST_BRLABEL) { \
         COMPUTE_DISP(ins); \
         predict = (disp != 0) ? 1 : 0; \
         g_assert (sparc_is_imm19 (disp)); \
-               sparc_branchp (code, (annul), (cond), sparc_icc_short, (predict), disp); \
+               sparc_branchp (code, (annul), (cond), DEFAULT_ICC, (predict), disp); \
         if (filldelay) sparc_nop (code); \
     } while (0)
 
@@ -1052,7 +1480,7 @@ if (ins->flags & MONO_INST_BRLABEL) { \
            gint32 disp; \
         COMPUTE_DISP(ins); \
                g_assert (sparc_is_imm22 (disp)); \
-               sparc_ ## bop (code, (annul), (predict), ins->sreg2, disp); \
+               sparc_ ## bop (code, (annul), (predict), ins->sreg1, disp); \
         if (filldelay) sparc_nop (code); \
     } while (0)
 
@@ -1060,18 +1488,27 @@ if (ins->flags & MONO_INST_BRLABEL) { \
 /*
  * We put the exception throwing code out-of-line, at the end of the method
  */
-#define EMIT_COND_SYSTEM_EXCEPTION(ins,cond,sexc_name) do {     \
+#define EMIT_COND_SYSTEM_EXCEPTION_GENERAL(ins,cond,sexc_name,filldelay,icc) do {     \
                mono_add_patch_info (cfg, (guint8*)(code) - (cfg)->native_code,   \
                                    MONO_PATCH_INFO_EXC, sexc_name);  \
-        if (sparcv9) { \
-           sparc_branchp (code, 0, (cond), sparc_icc_short, 0, 0); \
+        if (sparcv9 && ((icc) != sparc_icc_short)) {          \
+           sparc_branchp (code, 0, (cond), (icc), 0, 0); \
         } \
         else { \
-                       sparc_branch (code, 1, cond, 0);     \
+                       sparc_branch (code, 0, cond, 0);     \
         } \
-        sparc_nop (code);     \
+        if (filldelay) sparc_nop (code);     \
        } while (0); 
 
+#define EMIT_COND_SYSTEM_EXCEPTION(ins,cond,sexc_name) EMIT_COND_SYSTEM_EXCEPTION_GENERAL(ins,cond,sexc_name,TRUE,DEFAULT_ICC)
+
+#define EMIT_COND_SYSTEM_EXCEPTION_BPR(ins,bop,sexc_name) do { \
+               mono_add_patch_info (cfg, (guint8*)(code) - (cfg)->native_code,   \
+                                   MONO_PATCH_INFO_EXC, sexc_name);  \
+               sparc_ ## bop (code, FALSE, FALSE, ins->sreg1, 0); \
+        sparc_nop (code);    \
+} while (0);
+
 #define EMIT_ALU_IMM(ins,op,setcc) do { \
                        if (sparc_is_imm13 ((ins)->inst_imm)) \
                                sparc_ ## op ## _imm (code, (setcc), (ins)->sreg1, ins->inst_imm, (ins)->dreg); \
@@ -1100,8 +1537,8 @@ if (ins->flags & MONO_INST_BRLABEL) { \
                                sreg = sparc_o7; \
                        } \
                        if (!sparc_is_imm13 (ins->inst_offset)) { \
-                               sparc_set (code, ins->inst_offset, sparc_g1); \
-                               sparc_ ## op (code, sreg, ins->inst_destbasereg, sparc_g1); \
+                               sparc_set (code, ins->inst_offset, GP_SCRATCH_REG); \
+                               sparc_ ## op (code, sreg, ins->inst_destbasereg, GP_SCRATCH_REG); \
                        } \
                        else \
                                sparc_ ## op ## _imm (code, sreg, ins->inst_destbasereg, ins->inst_offset); \
@@ -1116,14 +1553,59 @@ if (ins->flags & MONO_INST_BRLABEL) { \
                                sparc_ ## op ## _imm (code, ins->sreg1, ins->inst_destbasereg, ins->inst_offset); \
                                                                                                                                                                                 } while (0);
 
-static void
-peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
+#define EMIT_CALL() do { \
+    if (v64) { \
+        sparc_set_template (code, sparc_o7); \
+        sparc_jmpl (code, sparc_o7, sparc_g0, sparc_o7); \
+    } \
+    else { \
+        sparc_call_simple (code, 0); \
+    } \
+    sparc_nop (code); \
+} while (0);
+
+/*
+ * A call template is 7 instructions long, so we want to avoid it if possible.
+ */
+static guint32*
+emit_call (MonoCompile *cfg, guint32 *code, guint32 patch_type, gconstpointer data)
 {
-       MonoInst *ins, *last_ins = NULL;
-       ins = bb->code;
+       gpointer target;
+
+       /* FIXME: This only works if the target method is already compiled */
+       if (0 && v64 && !cfg->compile_aot) {
+               MonoJumpInfo patch_info;
+
+               patch_info.type = patch_type;
+               patch_info.data.target = data;
+
+               target = mono_resolve_patch_target (cfg->method, cfg->domain, NULL, &patch_info, FALSE);
+
+               /* FIXME: Add optimizations if the target is close enough */
+               sparc_set (code, target, sparc_o7);
+               sparc_jmpl (code, sparc_o7, sparc_g0, sparc_o7);
+               sparc_nop (code);
+       }
+       else {
+               mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, patch_type, data);
+               EMIT_CALL ();
+       }
+       
+       return code;
+}
+
+void
+mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
+{
+}
 
-       while (ins) {
+void
+mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
+{
+       MonoInst *ins, *n, *last_ins = NULL;
+       ins = bb->code;
 
+       MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
                switch (ins->opcode) {
                case OP_MUL_IMM: 
                        /* remove unnecessary multiplication with 1 */
@@ -1131,12 +1613,12 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                                if (ins->dreg != ins->sreg1) {
                                        ins->opcode = OP_MOVE;
                                } else {
-                                       last_ins->next = ins->next;                             
-                                       ins = ins->next;                                
+                                       MONO_DELETE_INS (bb, ins);
                                        continue;
                                }
                        }
                        break;
+#ifndef SPARCV9
                case OP_LOAD_MEMBASE:
                case OP_LOADI4_MEMBASE:
                        /* 
@@ -1148,8 +1630,7 @@ 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_DELETE_INS (bb, ins);
                                        continue;
                                } else {
                                        //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
@@ -1172,8 +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_DELETE_INS (bb, ins);
                                        continue;
                                } else {
                                        ins->opcode = OP_MOVE;
@@ -1201,14 +1681,13 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
 #endif
                        }
                        break;
-               case OP_LOADU1_MEMBASE:
+#endif
                case OP_LOADI1_MEMBASE:
                        if (last_ins && (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
                                        ins->inst_basereg == last_ins->inst_destbasereg &&
                                        ins->inst_offset == last_ins->inst_offset) {
                                if (ins->dreg == last_ins->sreg1) {
-                                       last_ins->next = ins->next;                             
-                                       ins = ins->next;                                
+                                       MONO_DELETE_INS (bb, ins);
                                        continue;
                                } else {
                                        //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
@@ -1217,14 +1696,12 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                                }
                        }
                        break;
-               case OP_LOADU2_MEMBASE:
                case OP_LOADI2_MEMBASE:
                        if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
                                        ins->inst_basereg == last_ins->inst_destbasereg &&
                                        ins->inst_offset == last_ins->inst_offset) {
                                if (ins->dreg == last_ins->sreg1) {
-                                       last_ins->next = ins->next;                             
-                                       ins = ins->next;                                
+                                       MONO_DELETE_INS (bb, ins);
                                        continue;
                                } else {
                                        //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
@@ -1248,888 +1725,125 @@ peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
                                if (sparcv9) {
                                        last_ins->opcode = OP_STOREI8_MEMBASE_IMM;
                                        last_ins->inst_offset = ins->inst_offset;
-                                       last_ins->next = ins->next;                             
-                                       ins = ins->next;
+                                       MONO_DELETE_INS (bb, ins);
                                        continue;
                                }
                        }
                        break;
-               case CEE_BEQ:
-               case CEE_BNE_UN:
-               case CEE_BLT:
-               case CEE_BGT:
-               case CEE_BGE:
-               case CEE_BLE:
+               case OP_IBEQ:
+               case OP_IBNE_UN:
+               case OP_IBLT:
+               case OP_IBGT:
+               case OP_IBGE:
+               case OP_IBLE:
+               case OP_COND_EXC_EQ:
+               case OP_COND_EXC_GE:
+               case OP_COND_EXC_GT:
+               case OP_COND_EXC_LE:
+               case OP_COND_EXC_LT:
+               case OP_COND_EXC_NE_UN:
                        /*
                         * Convert compare with zero+branch to BRcc
                         */
                        /* 
                         * This only works in 64 bit mode, since it examines all 64
                         * bits of the register.
+                        * Only do this if the method is small since BPr only has a 16bit
+                        * displacement.
                         */
-                       if (0 && sparcv9 && last_ins && 
+                       if (v64 && (mono_method_get_header (cfg->method)->code_size < 10000) && last_ins && 
                                (last_ins->opcode == OP_COMPARE_IMM) &&
                                (last_ins->inst_imm == 0)) {
-                               MonoInst *next = ins->next;
                                switch (ins->opcode) {
-                               case CEE_BEQ:
+                               case OP_IBEQ:
                                        ins->opcode = OP_SPARC_BRZ;
                                        break;
-                               case CEE_BNE_UN:
+                               case OP_IBNE_UN:
                                        ins->opcode = OP_SPARC_BRNZ;
                                        break;
-                               case CEE_BLT:
+                               case OP_IBLT:
                                        ins->opcode = OP_SPARC_BRLZ;
                                        break;
-                               case CEE_BGT:
+                               case OP_IBGT:
                                        ins->opcode = OP_SPARC_BRGZ;
                                        break;
-                               case CEE_BGE:
+                               case OP_IBGE:
                                        ins->opcode = OP_SPARC_BRGEZ;
                                        break;
-                               case CEE_BLE:
+                               case OP_IBLE:
                                        ins->opcode = OP_SPARC_BRLEZ;
                                        break;
+                               case OP_COND_EXC_EQ:
+                                       ins->opcode = OP_SPARC_COND_EXC_EQZ;
+                                       break;
+                               case OP_COND_EXC_GE:
+                                       ins->opcode = OP_SPARC_COND_EXC_GEZ;
+                                       break;
+                               case OP_COND_EXC_GT:
+                                       ins->opcode = OP_SPARC_COND_EXC_GTZ;
+                                       break;
+                               case OP_COND_EXC_LE:
+                                       ins->opcode = OP_SPARC_COND_EXC_LEZ;
+                                       break;
+                               case OP_COND_EXC_LT:
+                                       ins->opcode = OP_SPARC_COND_EXC_LTZ;
+                                       break;
+                               case OP_COND_EXC_NE_UN:
+                                       ins->opcode = OP_SPARC_COND_EXC_NEZ;
+                                       break;
                                default:
                                        g_assert_not_reached ();
                                }
-                               ins->sreg2 = last_ins->sreg1;
+                               ins->sreg1 = last_ins->sreg1;
                                *last_ins = *ins;
-                               last_ins->next = next;
-                               ins = next;
+                               MONO_DELETE_INS (bb, ins);
                                continue;
                        }
                        break;
-               case CEE_CONV_I4:
-               case CEE_CONV_U4:
                case OP_MOVE:
                        /* 
                         * OP_MOVE reg, reg 
                         */
                        if (ins->dreg == ins->sreg1) {
-                               if (last_ins)
-                                       last_ins->next = ins->next;                             
-                               ins = ins->next;
+                               MONO_DELETE_INS (bb, ins);
                                continue;
                        }
                        /* 
                         * OP_MOVE sreg, dreg 
-                        * OP_MOVE dreg, sreg
-                        */
-                       if (last_ins && last_ins->opcode == OP_MOVE &&
-                           ins->sreg1 == last_ins->dreg &&
-                           ins->dreg == last_ins->sreg1) {
-                               last_ins->next = ins->next;                             
-                               ins = ins->next;                                
-                               continue;
-                       }
-                       break;
-               }
-               last_ins = ins;
-               ins = ins->next;
-       }
-       bb->last_ins = last_ins;
-}
-
-/* Parameters used by the register allocator */
-
-/* Use %l4..%l7 as local registers */
-#define ARCH_CALLER_REGS (0xf0<<16)
-/* Use %f2..%f30 as the double precision floating point local registers */
-#define ARCH_CALLER_FREGS (0x55555554)
-
-#undef DEBUG
-#define DEBUG(a) if (cfg->verbose_level > 1) a
-//#define DEBUG(a)
-#define reg_is_freeable(r) ((1 << (r)) & ARCH_CALLER_REGS)
-#define freg_is_freeable(r) (((1) << (r)) & ARCH_CALLER_FREGS)
-
-typedef struct {
-       int born_in;
-       int killed_in;
-       int last_use;
-       int prev_use;
-} RegTrack;
-
-static const char*const * ins_spec = sparc_desc;
-
-static void
-print_ins (int i, MonoInst *ins)
-{
-       const char *spec = ins_spec [ins->opcode];
-       g_print ("\t%-2d %s", i, mono_inst_name (ins->opcode));
-       if (spec [MONO_INST_DEST]) {
-               if (ins->dreg >= MONO_MAX_IREGS)
-                       g_print (" R%d <-", ins->dreg);
-               else
-                       if (spec [MONO_INST_DEST] == 'b')
-                               g_print (" [%s + 0x%x] <-", mono_arch_regname (ins->dreg), ins->inst_offset);
-               else
-                       g_print (" %s <-", mono_arch_regname (ins->dreg));
-       }
-       if (spec [MONO_INST_SRC1]) {
-               if (ins->sreg1 >= MONO_MAX_IREGS)
-                       g_print (" R%d", ins->sreg1);
-               else
-                       if (spec [MONO_INST_SRC1] == 'b')
-                               g_print (" [%s + 0x%x]", mono_arch_regname (ins->sreg1), ins->inst_offset);
-               else
-                       g_print (" %s", mono_arch_regname (ins->sreg1));
-       }
-       if (spec [MONO_INST_SRC2]) {
-               if (ins->sreg2 >= MONO_MAX_IREGS)
-                       g_print (" R%d", ins->sreg2);
-               else
-                       g_print (" %s", mono_arch_regname (ins->sreg2));
-       }
-       if (spec [MONO_INST_CLOB])
-               g_print (" clobbers: %c", spec [MONO_INST_CLOB]);
-       g_print ("\n");
-}
-
-static void
-print_regtrack (RegTrack *t, int num)
-{
-       int i;
-       char buf [32];
-       const char *r;
-       
-       for (i = 0; i < num; ++i) {
-               if (!t [i].born_in)
-                       continue;
-               if (i >= MONO_MAX_IREGS) {
-                       g_snprintf (buf, sizeof(buf), "R%d", i);
-                       r = buf;
-               } else
-                       r = mono_arch_regname (i);
-               g_print ("liveness: %s [%d - %d]\n", r, t [i].born_in, t[i].last_use);
-       }
-}
-
-typedef struct InstList InstList;
-
-struct InstList {
-       InstList *prev;
-       InstList *next;
-       MonoInst *data;
-};
-
-static inline InstList*
-inst_list_prepend (MonoMemPool *pool, InstList *list, MonoInst *data)
-{
-       InstList *item = mono_mempool_alloc (pool, sizeof (InstList));
-       item->data = data;
-       item->prev = NULL;
-       item->next = list;
-       if (list)
-               list->prev = item;
-       return item;
-}
-
-#define STACK_OFFSETS_POSITIVE
-
-/*
- * returns the offset used by spillvar. It allocates a new
- * spill variable if necessary.
- */
-static int
-mono_spillvar_offset (MonoCompile *cfg, int spillvar)
-{
-       MonoSpillInfo **si, *info;
-       int i = 0;
-
-       si = &cfg->spill_info; 
-       
-       while (i <= spillvar) {
-
-               if (!*si) {
-                       *si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
-                       info->next = NULL;
-                       cfg->stack_offset += sizeof (gpointer);
-                       info->offset = - cfg->stack_offset;
-               }
-
-               if (i == spillvar)
-                       return (*si)->offset;
-
-               i++;
-               si = &(*si)->next;
-       }
-
-       g_assert_not_reached ();
-       return 0;
-}
-
-static int
-mono_spillvar_offset_float (MonoCompile *cfg, int spillvar)
-{
-       MonoSpillInfo **si, *info;
-       int i = 0;
-
-       si = &cfg->spill_info_float; 
-       
-       while (i <= spillvar) {
-
-               if (!*si) {
-                       *si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
-                       info->next = NULL;
-                       cfg->stack_offset += sizeof (double);
-                       cfg->stack_offset = ALIGN_TO (cfg->stack_offset, 8);
-                       info->offset = - cfg->stack_offset;
-               }
-
-               if (i == spillvar)
-                       return (*si)->offset;
-
-               i++;
-               si = &(*si)->next;
-       }
-
-       g_assert_not_reached ();
-       return 0;
-}
-
-/*
- * Force the spilling of the variable in the symbolic register 'reg'.
- */
-static int
-get_register_force_spilling (MonoCompile *cfg, InstList *item, MonoInst *ins, int reg)
-{
-       MonoInst *load;
-       int i, sel, spill;
-       
-       sel = cfg->rs->iassign [reg];
-       /*i = cfg->rs->isymbolic [sel];
-       g_assert (i == reg);*/
-       i = reg;
-       spill = ++cfg->spill_count;
-       cfg->rs->iassign [i] = -spill - 1;
-       mono_regstate_free_int (cfg->rs, sel);
-       /* we need to create a spill var and insert a load to sel after the current instruction */
-       MONO_INST_NEW (cfg, load, OP_LOAD_MEMBASE);
-       load->dreg = sel;
-       load->inst_basereg = cfg->frame_reg;
-       load->inst_offset = mono_spillvar_offset (cfg, spill);
-       if (item->prev) {
-               while (ins->next != item->prev->data)
-                       ins = ins->next;
-       }
-       load->next = ins->next;
-       ins->next = load;
-       DEBUG (g_print ("SPILLED LOAD (%d at 0x%08x(%%sp)) R%d (freed %s)\n", spill, load->inst_offset, i, mono_arch_regname (sel)));
-       i = mono_regstate_alloc_int (cfg->rs, 1 << sel);
-       g_assert (i == sel);
-
-       return sel;
-}
-
-static int
-get_register_spilling (MonoCompile *cfg, InstList *item, MonoInst *ins, guint32 regmask, int reg)
-{
-       MonoInst *load;
-       int i, sel, spill;
-
-       DEBUG (g_print ("start regmask to assign R%d: 0x%08x (R%d <- R%d R%d)\n", reg, regmask, ins->dreg, ins->sreg1, ins->sreg2));
-       /* exclude the registers in the current instruction */
-       if (reg != ins->sreg1 && (reg_is_freeable (ins->sreg1) || (ins->sreg1 >= MONO_MAX_IREGS && cfg->rs->iassign [ins->sreg1] >= 0))) {
-               if (ins->sreg1 >= MONO_MAX_IREGS)
-                       regmask &= ~ (1 << cfg->rs->iassign [ins->sreg1]);
-               else
-                       regmask &= ~ (1 << ins->sreg1);
-               DEBUG (g_print ("excluding sreg1 %s\n", mono_arch_regname (ins->sreg1)));
-       }
-       if (reg != ins->sreg2 && (reg_is_freeable (ins->sreg2) || (ins->sreg2 >= MONO_MAX_IREGS && cfg->rs->iassign [ins->sreg2] >= 0))) {
-               if (ins->sreg2 >= MONO_MAX_IREGS)
-                       regmask &= ~ (1 << cfg->rs->iassign [ins->sreg2]);
-               else
-                       regmask &= ~ (1 << ins->sreg2);
-               DEBUG (g_print ("excluding sreg2 %s %d\n", mono_arch_regname (ins->sreg2), ins->sreg2));
-       }
-       if (reg != ins->dreg && reg_is_freeable (ins->dreg)) {
-               regmask &= ~ (1 << ins->dreg);
-               DEBUG (g_print ("excluding dreg %s\n", mono_arch_regname (ins->dreg)));
-       }
-
-       DEBUG (g_print ("available regmask: 0x%08x\n", regmask));
-       g_assert (regmask); /* need at least a register we can free */
-       sel = -1;
-       /* we should track prev_use and spill the register that's farther */
-       for (i = 0; i < MONO_MAX_IREGS; ++i) {
-               if (regmask & (1 << i)) {
-                       sel = i;
-                       DEBUG (g_print ("selected register %s has assignment %d\n", mono_arch_regname (sel), cfg->rs->iassign [sel]));
-                       break;
-               }
-       }
-       i = cfg->rs->isymbolic [sel];
-       spill = ++cfg->spill_count;
-       cfg->rs->iassign [i] = -spill - 1;
-       mono_regstate_free_int (cfg->rs, sel);
-       /* we need to create a spill var and insert a load to sel after the current instruction */
-       MONO_INST_NEW (cfg, load, OP_LOAD_MEMBASE);
-       load->dreg = sel;
-       load->inst_basereg = cfg->frame_reg;
-       load->inst_offset = mono_spillvar_offset (cfg, spill);
-       if (item->prev) {
-               while (ins->next != item->prev->data)
-                       ins = ins->next;
-       }
-       load->next = ins->next;
-       ins->next = load;
-       DEBUG (g_print ("SPILLED LOAD (%d at 0x%08x(%%sp)) R%d (freed %s)\n", spill, load->inst_offset, i, mono_arch_regname (sel)));
-       i = mono_regstate_alloc_int (cfg->rs, 1 << sel);
-       g_assert (i == sel);
-       
-       return sel;
-}
-
-static int
-get_float_register_spilling (MonoCompile *cfg, InstList *item, MonoInst *ins, guint32 regmask, int reg)
-{
-       MonoInst *load;
-       int i, sel, spill;
-
-       DEBUG (g_print ("start regmask to assign R%d: 0x%08x (R%d <- R%d R%d)\n", reg, regmask, ins->dreg, ins->sreg1, ins->sreg2));
-       /* exclude the registers in the current instruction */
-       if (reg != ins->sreg1 && (freg_is_freeable (ins->sreg1) || (ins->sreg1 >= MONO_MAX_FREGS && cfg->rs->fassign [ins->sreg1] >= 0))) {
-               if (ins->sreg1 >= MONO_MAX_FREGS)
-                       regmask &= ~ (1 << cfg->rs->fassign [ins->sreg1]);
-               else
-                       regmask &= ~ (1 << ins->sreg1);
-               DEBUG (g_print ("excluding sreg1 %s\n", mono_arch_regname (ins->sreg1)));
-       }
-       if (reg != ins->sreg2 && (freg_is_freeable (ins->sreg2) || (ins->sreg2 >= MONO_MAX_FREGS && cfg->rs->fassign [ins->sreg2] >= 0))) {
-               if (ins->sreg2 >= MONO_MAX_FREGS)
-                       regmask &= ~ (1 << cfg->rs->fassign [ins->sreg2]);
-               else
-                       regmask &= ~ (1 << ins->sreg2);
-               DEBUG (g_print ("excluding sreg2 %s %d\n", mono_arch_regname (ins->sreg2), ins->sreg2));
-       }
-       if (reg != ins->dreg && freg_is_freeable (ins->dreg)) {
-               regmask &= ~ (1 << ins->dreg);
-               DEBUG (g_print ("excluding dreg %s\n", mono_arch_regname (ins->dreg)));
-       }
-
-       DEBUG (g_print ("available regmask: 0x%08x\n", regmask));
-       g_assert (regmask); /* need at least a register we can free */
-       sel = -1;
-       /* we should track prev_use and spill the register that's farther */
-       for (i = 0; i < MONO_MAX_FREGS; ++i) {
-               if (regmask & (1 << i)) {
-                       sel = i;
-                       DEBUG (g_print ("selected register %s has assignment %d\n", mono_arch_regname (sel), cfg->rs->fassign [sel]));
-                       break;
-               }
-       }
-       i = cfg->rs->fsymbolic [sel];
-       spill = ++cfg->spill_count;
-       cfg->rs->fassign [i] = -spill - 1;
-       mono_regstate_free_float(cfg->rs, sel);
-       /* we need to create a spill var and insert a load to sel after the current instruction */
-       MONO_INST_NEW (cfg, load, OP_LOADR8_MEMBASE);
-       load->dreg = sel;
-       load->inst_basereg = cfg->frame_reg;
-       load->inst_offset = mono_spillvar_offset_float (cfg, spill);
-       if (item->prev) {
-               while (ins->next != item->prev->data)
-                       ins = ins->next;
-       }
-       load->next = ins->next;
-       ins->next = load;
-       DEBUG (g_print ("SPILLED LOAD (%d at 0x%08x(%%sp)) R%d (freed %s)\n", spill, load->inst_offset, i, mono_arch_regname (sel)));
-       i = mono_regstate_alloc_float (cfg->rs, 1 << sel);
-       g_assert (i == sel);
-       
-       return sel;
-}
-
-static MonoInst*
-create_copy_ins (MonoCompile *cfg, int dest, int src, MonoInst *ins)
-{
-       MonoInst *copy;
-       MONO_INST_NEW (cfg, copy, OP_MOVE);
-       copy->dreg = dest;
-       copy->sreg1 = src;
-       if (ins) {
-               copy->next = ins->next;
-               ins->next = copy;
-       }
-       DEBUG (g_print ("\tforced copy from %s to %s\n", mono_arch_regname (src), mono_arch_regname (dest)));
-       return copy;
-}
-
-static MonoInst*
-create_copy_ins_float (MonoCompile *cfg, int dest, int src, MonoInst *ins)
-{
-       MonoInst *copy;
-       MONO_INST_NEW (cfg, copy, OP_FMOVE);
-       copy->dreg = dest;
-       copy->sreg1 = src;
-       if (ins) {
-               copy->next = ins->next;
-               ins->next = copy;
-       }
-       DEBUG (g_print ("\tforced copy from %s to %s\n", mono_arch_regname (src), mono_arch_regname (dest)));
-       return copy;
-}
-
-static MonoInst*
-create_spilled_store (MonoCompile *cfg, int spill, int reg, int prev_reg, MonoInst *ins)
-{
-       MonoInst *store;
-       MONO_INST_NEW (cfg, store, OP_STORE_MEMBASE_REG);
-       store->sreg1 = reg;
-       store->inst_destbasereg = cfg->frame_reg;
-       store->inst_offset = mono_spillvar_offset (cfg, spill);
-       if (ins) {
-               store->next = ins->next;
-               ins->next = store;
-       }
-       DEBUG (g_print ("SPILLED STORE (%d at 0x%08x(%%sp)) R%d (from %s)\n", spill, store->inst_offset, prev_reg, mono_arch_regname (reg)));
-       return store;
-}
-
-static MonoInst*
-create_spilled_store_float (MonoCompile *cfg, int spill, int reg, int prev_reg, MonoInst *ins)
-{
-       MonoInst *store;
-       MONO_INST_NEW (cfg, store, OP_STORER8_MEMBASE_REG);
-       store->sreg1 = reg;
-       store->inst_destbasereg = cfg->frame_reg;
-       store->inst_offset = mono_spillvar_offset_float (cfg, spill);
-       if (ins) {
-               store->next = ins->next;
-               ins->next = store;
-       }
-       DEBUG (g_print ("SPILLED STORE (%d at 0x%08x(%%sp)) R%d (from %s)\n", spill, store->inst_offset, prev_reg, mono_arch_regname (reg)));
-       return store;
-}
-
-static void
-insert_before_ins (MonoInst *ins, InstList *item, MonoInst* to_insert)
-{
-       MonoInst *prev;
-       g_assert (item->next);
-       prev = item->next->data;
-
-       while (prev->next != ins)
-               prev = prev->next;
-       to_insert->next = ins;
-       prev->next = to_insert;
-       /* 
-        * needed otherwise in the next instruction we can add an ins to the 
-        * end and that would get past this instruction.
-        */
-       item->data = to_insert; 
-}
-
-static int
-alloc_int_reg (MonoCompile *cfg, InstList *curinst, MonoInst *ins, int sym_reg, guint32 allow_mask)
-{
-       int val = cfg->rs->iassign [sym_reg];
-       if (val < 0) {
-               int spill = 0;
-               if (val < -1) {
-                       /* the register gets spilled after this inst */
-                       spill = -val -1;
-               }
-               val = mono_regstate_alloc_int (cfg->rs, allow_mask);
-               if (val < 0)
-                       val = get_register_spilling (cfg, curinst, ins, allow_mask, sym_reg);
-               cfg->rs->iassign [sym_reg] = val;
-               /* add option to store before the instruction for src registers */
-               if (spill)
-                       create_spilled_store (cfg, spill, val, sym_reg, ins);
-       }
-       cfg->rs->isymbolic [val] = sym_reg;
-       return val;
-}
-
-/* FIXME: Strange loads from the stack in basic-float.cs:test_2_rem */
-
-/*
- * Local register allocation.
- * We first scan the list of instructions and we save the liveness info of
- * each register (when the register is first used, when it's value is set etc.).
- * We also reverse the list of instructions (in the InstList list) because assigning
- * registers backwards allows for more tricks to be used.
- */
-void
-mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
-{
-       MonoInst *ins;
-       MonoRegState *rs = cfg->rs;
-       int i, val;
-       RegTrack *reginfo, *reginfof;
-       RegTrack *reginfo1, *reginfo2, *reginfod;
-       InstList *tmp, *reversed = NULL;
-       const char *spec;
-       guint32 src1_mask, src2_mask, dest_mask;
-       guint32 cur_iregs, cur_fregs;
-
-       /* FIXME: Use caller saved regs and %i1-%2 for allocation */
-
-       if (!bb->code)
-               return;
-       rs->next_vireg = bb->max_ireg;
-       rs->next_vfreg = bb->max_freg;
-       mono_regstate_assign (rs);
-       reginfo = mono_mempool_alloc0 (cfg->mempool, sizeof (RegTrack) * rs->next_vireg);
-       reginfof = mono_mempool_alloc0 (cfg->mempool, sizeof (RegTrack) * rs->next_vfreg);
-       rs->ifree_mask = ARCH_CALLER_REGS;
-       rs->ffree_mask = ARCH_CALLER_FREGS;
-
-       ins = bb->code;
-       i = 1;
-       DEBUG (g_print ("LOCAL regalloc: basic block: %d\n", bb->block_num));
-       /* forward pass on the instructions to collect register liveness info */
-       while (ins) {
-               spec = ins_spec [ins->opcode];
-               if (!spec) {
-                       mono_print_tree_nl (ins);
-                       g_assert (spec);
-               }
-               DEBUG (print_ins (i, ins));
-
-               if (spec [MONO_INST_SRC1]) {
-                       if (spec [MONO_INST_SRC1] == 'f')
-                               reginfo1 = reginfof;
-                       else
-                               reginfo1 = reginfo;
-                       reginfo1 [ins->sreg1].prev_use = reginfo1 [ins->sreg1].last_use;
-                       reginfo1 [ins->sreg1].last_use = i;
-               } else {
-                       ins->sreg1 = -1;
-               }
-               if (spec [MONO_INST_SRC2]) {
-                       if (spec [MONO_INST_SRC2] == 'f')
-                               reginfo2 = reginfof;
-                       else
-                               reginfo2 = reginfo;
-                       reginfo2 [ins->sreg2].prev_use = reginfo2 [ins->sreg2].last_use;
-                       reginfo2 [ins->sreg2].last_use = i;
-               } else {
-                       ins->sreg2 = -1;
-               }
-               if (spec [MONO_INST_DEST]) {
-                       if (spec [MONO_INST_DEST] == 'f')
-                               reginfod = reginfof;
-                       else
-                               reginfod = reginfo;
-                       if (spec [MONO_INST_DEST] != 'b') /* it's not just a base register */
-                               reginfod [ins->dreg].killed_in = i;
-                       reginfod [ins->dreg].prev_use = reginfod [ins->dreg].last_use;
-                       reginfod [ins->dreg].last_use = i;
-                       if (reginfod [ins->dreg].born_in == 0 || reginfod [ins->dreg].born_in > i)
-                               reginfod [ins->dreg].born_in = i;
-                       if (spec [MONO_INST_DEST] == 'l') {
-                               /* result in eax:edx, the virtual register is allocated sequentially */
-                               reginfod [ins->dreg + 1].prev_use = reginfod [ins->dreg + 1].last_use;
-                               reginfod [ins->dreg + 1].last_use = i;
-                               if (reginfod [ins->dreg + 1].born_in == 0 || reginfod [ins->dreg + 1].born_in > i)
-                                       reginfod [ins->dreg + 1].born_in = i;
-                       }
-               } else {
-                       ins->dreg = -1;
-               }
-               reversed = inst_list_prepend (cfg->mempool, reversed, ins);
-               ++i;
-               ins = ins->next;
-       }
-
-       cur_iregs = ARCH_CALLER_REGS;
-       cur_fregs = ARCH_CALLER_FREGS;
-
-       DEBUG (print_regtrack (reginfo, rs->next_vireg));
-       DEBUG (print_regtrack (reginfof, rs->next_vfreg));
-       tmp = reversed;
-       while (tmp) {
-               int prev_dreg, prev_sreg1, prev_sreg2;
-               --i;
-               ins = tmp->data;
-               spec = ins_spec [ins->opcode];
-               DEBUG (g_print ("processing:"));
-               DEBUG (print_ins (i, ins));
-
-               /* make the register available for allocation: FIXME add fp reg */
-               if (ins->opcode == OP_SETREG || ins->opcode == OP_SETREGIMM) {
-                       /* Dont free register which can't be allocated */
-                       if (reg_is_freeable (ins->dreg)) {
-                               cur_iregs |= 1 << ins->dreg;
-                               DEBUG (g_print ("adding %d to cur_iregs\n", ins->dreg));
-                       }
-               } else if (ins->opcode == OP_SETFREG) {
-                       if (freg_is_freeable (ins->dreg)) {
-                               cur_fregs |= 1 << ins->dreg;
-                               DEBUG (g_print ("adding %d to cur_fregs\n", ins->dreg));
-                       }
-               } else if (spec [MONO_INST_CLOB] == 'c') {
-                       MonoCallInst *cinst = (MonoCallInst*)ins;
-                       DEBUG (g_print ("excluding regs 0x%x from cur_iregs (0x%x)\n", cinst->used_iregs, cur_iregs));
-                       cur_iregs &= ~cinst->used_iregs;
-                       cur_fregs &= ~cinst->used_fregs;
-                       DEBUG (g_print ("available cur_iregs: 0x%x\n", cur_iregs));
-                       /* registers used by the calling convention are excluded from 
-                        * allocation: they will be selectively enabled when they are 
-                        * assigned by the special SETREG opcodes.
-                        */
-               }
-               dest_mask = src1_mask = src2_mask = cur_iregs;
-
-               /*
-                * DEST
-                */
-               /* update for use with FP regs... */
-               if (spec [MONO_INST_DEST] == 'f') {
-                       if (ins->dreg >= MONO_MAX_FREGS) {
-                               val = rs->fassign [ins->dreg];
-                               prev_dreg = ins->dreg;
-                               if (val < 0) {
-                                       int spill = 0;
-                                       if (val < -1) {
-                                               /* the register gets spilled after this inst */
-                                               spill = -val -1;
-                                       }
-                                       dest_mask = cur_fregs;
-                                       val = mono_regstate_alloc_float (rs, dest_mask);
-                                       if (val < 0)
-                                               val = get_float_register_spilling (cfg, tmp, ins, dest_mask, ins->dreg);
-                                       rs->fassign [ins->dreg] = val;
-                                       if (spill)
-                                               create_spilled_store_float (cfg, spill, val, prev_dreg, ins);
-                               }
-                               DEBUG (g_print ("\tassigned dreg %s to dest R%d\n", mono_arch_regname (val), ins->dreg));
-                               rs->fsymbolic [val] = prev_dreg;
-                               ins->dreg = val;
-                       } else {
-                               prev_dreg = -1;
-                       }
-                       if (freg_is_freeable (ins->dreg) && prev_dreg >= 0 && (reginfo [prev_dreg].born_in >= i || !(cur_fregs & (1 << ins->dreg)))) {
-                               DEBUG (g_print ("\tfreeable %s (R%d) (born in %d)\n", mono_arch_regname (ins->dreg), prev_dreg, reginfo [prev_dreg].born_in));
-                               mono_regstate_free_float (rs, ins->dreg);
-                       }
-               } else if (ins->dreg >= MONO_MAX_IREGS) {
-                       val = rs->iassign [ins->dreg];
-                       prev_dreg = ins->dreg;
-                       if (val < 0) {
-                               int spill = 0;
-                               if (val < -1) {
-                                       /* the register gets spilled after this inst */
-                                       spill = -val -1;
-                               }
-                               val = mono_regstate_alloc_int (rs, dest_mask);
-                               if (val < 0)
-                                       val = get_register_spilling (cfg, tmp, ins, dest_mask, ins->dreg);
-                               rs->iassign [ins->dreg] = val;
-                               if (spill)
-                                       create_spilled_store (cfg, spill, val, prev_dreg, ins);
-                       }
-                       DEBUG (g_print ("\tassigned dreg %s to dest R%d\n", mono_arch_regname (val), ins->dreg));
-                       rs->isymbolic [val] = prev_dreg;
-                       ins->dreg = val;
-                       if (spec [MONO_INST_DEST] == 'l') {
-                               int hreg = prev_dreg + 1;
-                               val = rs->iassign [hreg];
-                               if (val < 0) {
-                                       int spill = 0;
-                                       if (val < -1) {
-                                               /* the register gets spilled after this inst */
-                                               spill = -val -1;
-                                       }
-                                       /* The second register must be a pair of the first */
-                                       dest_mask = 1 << (rs->iassign [prev_dreg] + 1);
-                                       val = mono_regstate_alloc_int (rs, dest_mask);
-                                       if (val < 0)
-                                               val = get_register_spilling (cfg, tmp, ins, dest_mask, hreg);
-                                       rs->iassign [hreg] = val;
-                                       if (spill)
-                                               create_spilled_store (cfg, spill, val, hreg, ins);
-                               }
-                               else {
-                                       /* The second register must be a pair of the first */
-                                       if (val != rs->iassign [prev_dreg] + 1) {
-                                               dest_mask = 1 << (rs->iassign [prev_dreg] + 1);
-
-                                               val = mono_regstate_alloc_int (rs, dest_mask);
-                                               if (val < 0)
-                                                       val = get_register_spilling (cfg, tmp, ins, dest_mask, hreg);
-
-                                               create_copy_ins (cfg, rs->iassign [hreg], val, ins);
-
-                                               rs->iassign [hreg] = val;
-                                       }
-                               }                                       
-
-                               DEBUG (g_print ("\tassigned hreg %s to dest R%d\n", mono_arch_regname (val), hreg));
-                               rs->isymbolic [val] = hreg;
-
-                               if (reg_is_freeable (val) && hreg >= 0 && (reginfo [hreg].born_in >= i && !(cur_iregs & (1 << val)))) {
-                                       DEBUG (g_print ("\tfreeable %s (R%d)\n", mono_arch_regname (val), hreg));
-                                       mono_regstate_free_int (rs, val);
-                               }
-                       }
-               } else {
-                       prev_dreg = -1;
-               }
-               if (spec [MONO_INST_DEST] != 'f' && reg_is_freeable (ins->dreg) && prev_dreg >= 0 && (reginfo [prev_dreg].born_in >= i)) {
-                       DEBUG (g_print ("\tfreeable %s (R%d) (born in %d)\n", mono_arch_regname (ins->dreg), prev_dreg, reginfo [prev_dreg].born_in));
-                       mono_regstate_free_int (rs, ins->dreg);
-               }
-
-               /**
-                * SRC1
-                */
-               if (spec [MONO_INST_SRC1] == 'f') {
-                       if (ins->sreg1 >= MONO_MAX_FREGS) {
-                               val = rs->fassign [ins->sreg1];
-                               prev_sreg1 = ins->sreg1;
-                               if (val < 0) {
-                                       int spill = 0;
-                                       if (val < -1) {
-                                               /* the register gets spilled after this inst */
-                                               spill = -val -1;
-                                       }
-                                       //g_assert (val == -1); /* source cannot be spilled */
-                                       src1_mask = cur_fregs;
-                                       val = mono_regstate_alloc_float (rs, src1_mask);
-                                       if (val < 0)
-                                               val = get_float_register_spilling (cfg, tmp, ins, src1_mask, ins->sreg1);
-                                       rs->fassign [ins->sreg1] = val;
-                                       DEBUG (g_print ("\tassigned sreg1 %s to R%d\n", mono_arch_regname (val), ins->sreg1));
-                                       if (spill) {
-                                               MonoInst *store = create_spilled_store_float (cfg, spill, val, prev_sreg1, NULL);
-                                               insert_before_ins (ins, tmp, store);
-                                       }
-                               }
-                               rs->fsymbolic [val] = prev_sreg1;
-                               ins->sreg1 = val;
-                       } else {
-                               prev_sreg1 = -1;
-                       }
-               } else if (ins->sreg1 >= MONO_MAX_IREGS) {
-                       val = rs->iassign [ins->sreg1];
-                       prev_sreg1 = ins->sreg1;
-                       if (val < 0) {
-                               int spill = 0;
-                               if (val < -1) {
-                                       /* the register gets spilled after this inst */
-                                       spill = -val -1;
-                               }
-                               if (0 && (ins->opcode == OP_MOVE) && reg_is_freeable (ins->dreg)) {
-                                       /* 
-                                        * small optimization: the dest register is already allocated
-                                        * but the src one is not: we can simply assign the same register
-                                        * here and peephole will get rid of the instruction later.
-                                        * This optimization may interfere with the clobbering handling:
-                                        * it removes a mov operation that will be added again to handle clobbering.
-                                        * There are also some other issues that should with make testjit.
-                                        */
-                                       mono_regstate_alloc_int (rs, 1 << ins->dreg);
-                                       val = rs->iassign [ins->sreg1] = ins->dreg;
-                                       //g_assert (val >= 0);
-                                       DEBUG (g_print ("\tfast assigned sreg1 %s to R%d\n", mono_arch_regname (val), ins->sreg1));
-                               } else {
-                                       //g_assert (val == -1); /* source cannot be spilled */
-                                       val = mono_regstate_alloc_int (rs, src1_mask);
-                                       if (val < 0)
-                                               val = get_register_spilling (cfg, tmp, ins, src1_mask, ins->sreg1);
-                                       rs->iassign [ins->sreg1] = val;
-                                       DEBUG (g_print ("\tassigned sreg1 %s to R%d\n", mono_arch_regname (val), ins->sreg1));
-                               }
-                               if (spill) {
-                                       MonoInst *store = create_spilled_store (cfg, spill, val, prev_sreg1, NULL);
-                                       insert_before_ins (ins, tmp, store);
-                               }
-                       }
-                       rs->isymbolic [val] = prev_sreg1;
-                       ins->sreg1 = val;
-               } else {
-                       prev_sreg1 = -1;
-               }
-
-               /*
-                * SRC2
-                */
-               if (spec [MONO_INST_SRC2] == 'f') {
-                       if (ins->sreg2 >= MONO_MAX_FREGS) {
-                               val = rs->fassign [ins->sreg2];
-                               prev_sreg2 = ins->sreg2;
-                               if (val < 0) {
-                                       int spill = 0;
-                                       if (val < -1) {
-                                               /* the register gets spilled after this inst */
-                                               spill = -val -1;
-                                       }
-                                       src2_mask = cur_fregs;
-                                       val = mono_regstate_alloc_float (rs, src2_mask);
-                                       if (val < 0)
-                                               val = get_float_register_spilling (cfg, tmp, ins, src2_mask, ins->sreg2);
-                                       rs->fassign [ins->sreg2] = val;
-                                       DEBUG (g_print ("\tassigned sreg2 %s to R%d\n", mono_arch_regname (val), ins->sreg2));
-                                       if (spill)
-                                               create_spilled_store_float (cfg, spill, val, prev_sreg2, ins);
-                               }
-                               rs->fsymbolic [val] = prev_sreg2;
-                               ins->sreg2 = val;
-                       } else {
-                               prev_sreg2 = -1;
-                       }
-               } else if (ins->sreg2 >= MONO_MAX_IREGS) {
-                       val = rs->iassign [ins->sreg2];
-                       prev_sreg2 = ins->sreg2;
-                       if (val < 0) {
-                               int spill = 0;
-                               if (val < -1) {
-                                       /* the register gets spilled after this inst */
-                                       spill = -val -1;
-                               }
-                               val = mono_regstate_alloc_int (rs, src2_mask);
-                               if (val < 0)
-                                       val = get_register_spilling (cfg, tmp, ins, src2_mask, ins->sreg2);
-                               rs->iassign [ins->sreg2] = val;
-                               DEBUG (g_print ("\tassigned sreg2 %s to R%d\n", mono_arch_regname (val), ins->sreg2));
-                               if (spill)
-                                       create_spilled_store (cfg, spill, val, prev_sreg2, ins);
-                       }
-                       rs->isymbolic [val] = prev_sreg2;
-                       ins->sreg2 = val;
-               } else {
-                       prev_sreg2 = -1;
-               }
-
-               if (spec [MONO_INST_CLOB] == 'c') {
-                       int j, s;
-                       guint32 clob_mask = ARCH_CALLER_REGS;
-                       for (j = 0; j < MONO_MAX_IREGS; ++j) {
-                               s = 1 << j;
-                               if ((clob_mask & s) && !(rs->ifree_mask & s) && j != ins->sreg1) {
-                                       //g_warning ("register %s busy at call site\n", mono_arch_regname (j));
-                               }
+                        * OP_MOVE dreg, sreg
+                        */
+                       if (last_ins && last_ins->opcode == OP_MOVE &&
+                           ins->sreg1 == last_ins->dreg &&
+                           ins->dreg == last_ins->sreg1) {
+                               MONO_DELETE_INS (bb, ins);
+                               continue;
                        }
+                       break;
                }
-               /*if (reg_is_freeable (ins->sreg1) && prev_sreg1 >= 0 && reginfo [prev_sreg1].born_in >= i) {
-                       DEBUG (g_print ("freeable %s\n", mono_arch_regname (ins->sreg1)));
-                       mono_regstate_free_int (rs, ins->sreg1);
-               }
-               if (reg_is_freeable (ins->sreg2) && prev_sreg2 >= 0 && reginfo [prev_sreg2].born_in >= i) {
-                       DEBUG (g_print ("freeable %s\n", mono_arch_regname (ins->sreg2)));
-                       mono_regstate_free_int (rs, ins->sreg2);
-               }*/
-               
-               //DEBUG (print_ins (i, ins));
-
-               tmp = tmp->next;
+               last_ins = ins;
+               ins = ins->next;
        }
+       bb->last_ins = last_ins;
+}
+
+void
+mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
+{
 }
 
+/* FIXME: Strange loads from the stack in basic-float.cs:test_2_rem */
+
 static void
-sparc_patch (guint8 *code, const guint8 *target)
+sparc_patch (guint32 *code, const gpointer target)
 {
-       guint32 ins = *(guint32*)code;
+       guint32 *c = code;
+       guint32 ins = *code;
        guint32 op = ins >> 30;
        guint32 op2 = (ins >> 22) & 0x7;
        guint32 rd = (ins >> 25) & 0x1f;
-       gint32 disp = (target - code) >> 2;
+       guint8* target8 = (guint8*)target;
+       gint64 disp = (target8 - (guint8*)code) >> 2;
+       int reg;
 
 //     g_print ("patching 0x%08x (0x%08x) to point to 0x%08x\n", code, ins, target);
 
@@ -2137,72 +1851,103 @@ sparc_patch (guint8 *code, const guint8 *target)
                if (!sparc_is_imm22 (disp))
                        NOT_IMPLEMENTED;
                /* Bicc */
-               *(guint32*)code = ((ins >> 22) << 22) | (disp & 0x3fffff);
+               *code = ((ins >> 22) << 22) | (disp & 0x3fffff);
        }
        else if ((op == 0) && (op2 == 1)) {
                if (!sparc_is_imm19 (disp))
                        NOT_IMPLEMENTED;
                /* BPcc */
-               *(guint32*)code = ((ins >> 19) << 19) | (disp & 0x7ffff);
+               *code = ((ins >> 19) << 19) | (disp & 0x7ffff);
        }
        else if ((op == 0) && (op2 == 3)) {
                if (!sparc_is_imm16 (disp))
                        NOT_IMPLEMENTED;
                /* BPr */
-               *(guint32*)code &= ~(0x180000 | 0x3fff);
-               *(guint32*)code |= ((disp << 21) & (0x180000)) | (disp & 0x3fff);
+               *code &= ~(0x180000 | 0x3fff);
+               *code |= ((disp << 21) & (0x180000)) | (disp & 0x3fff);
        }
        else if ((op == 0) && (op2 == 6)) {
                if (!sparc_is_imm22 (disp))
                        NOT_IMPLEMENTED;
                /* FBicc */
-               *(guint32*)code = ((ins >> 22) << 22) | (disp & 0x3fffff);
+               *code = ((ins >> 22) << 22) | (disp & 0x3fffff);
        }
        else if ((op == 0) && (op2 == 4)) {
-               guint32 ins2 = *(guint32*)(code + 4);
+               guint32 ins2 = code [1];
 
                if (((ins2 >> 30) == 2) && (((ins2 >> 19) & 0x3f) == 2)) {
-                       /* sethi followed by or */
-                       guint32 *p = (guint32*)code;
-                       sparc_set (p, target, rd);
-                       while (p <= (code + 4))
+                       /* sethi followed by or */                      
+                       guint32 *p = code;
+                       sparc_set (p, target8, rd);
+                       while (p <= (code + 1))
                                sparc_nop (p);
                }
                else if (ins2 == 0x01000000) {
                        /* sethi followed by nop */
-                       guint32 *p = (guint32*)code;
-                       sparc_set (p, target, rd);
-                       while (p <= (code + 4))
+                       guint32 *p = code;
+                       sparc_set (p, target8, rd);
+                       while (p <= (code + 1))
                                sparc_nop (p);
                }
                else if ((sparc_inst_op (ins2) == 3) && (sparc_inst_imm (ins2))) {
                        /* sethi followed by load/store */
-                       guint32 t = (guint32)target;
-                       *(guint32*)code &= ~(0x3fffff);
-                       *(guint32*)code |= (t >> 10);
-                       *(guint32*)(code + 4) &= ~(0x3ff);
-                       *(guint32*)(code + 4) |= (t & 0x3ff);
+#ifndef SPARCV9
+                       guint32 t = (guint32)target8;
+                       *code &= ~(0x3fffff);
+                       *code |= (t >> 10);
+                       *(code + 1) &= ~(0x3ff);
+                       *(code + 1) |= (t & 0x3ff);
+#endif
+               }
+               else if (v64 && 
+                                (sparc_inst_rd (ins) == sparc_g1) &&
+                                (sparc_inst_op (c [1]) == 0) && (sparc_inst_op2 (c [1]) == 4) &&
+                                (sparc_inst_op (c [2]) == 2) && (sparc_inst_op3 (c [2]) == 2) &&
+                                (sparc_inst_op (c [3]) == 2) && (sparc_inst_op3 (c [3]) == 2))
+               {
+                       /* sparc_set */
+                       guint32 *p = c;
+                       reg = sparc_inst_rd (c [1]);
+                       sparc_set (p, target8, reg);
+                       while (p < (c + 6))
+                               sparc_nop (p);
                }
                else if ((sparc_inst_op (ins2) == 2) && (sparc_inst_op3 (ins2) == 0x38) && 
                                 (sparc_inst_imm (ins2))) {
                        /* sethi followed by jmpl */
-                       guint32 t = (guint32)target;
-                       *(guint32*)code &= ~(0x3fffff);
-                       *(guint32*)code |= (t >> 10);
-                       *(guint32*)(code + 4) &= ~(0x3ff);
-                       *(guint32*)(code + 4) |= (t & 0x3ff);
+#ifndef SPARCV9
+                       guint32 t = (guint32)target8;
+                       *code &= ~(0x3fffff);
+                       *code |= (t >> 10);
+                       *(code + 1) &= ~(0x3ff);
+                       *(code + 1) |= (t & 0x3ff);
+#endif
                }
                else
                        NOT_IMPLEMENTED;
        }
        else if (op == 01) {
-               sparc_call_simple (code, target - code);
+               gint64 disp = (target8 - (guint8*)code) >> 2;
+
+               if (!sparc_is_imm30 (disp))
+                       NOT_IMPLEMENTED;
+               sparc_call_simple (code, target8 - (guint8*)code);
        }
        else if ((op == 2) && (sparc_inst_op3 (ins) == 0x2) && sparc_inst_imm (ins)) {
                /* mov imm, reg */
-               g_assert (sparc_is_imm13 (target));
-               *(guint32*)code &= ~(0x1fff);
-               *(guint32*)code |= (guint32)target;
+               g_assert (sparc_is_imm13 (target8));
+               *code &= ~(0x1fff);
+               *code |= (guint32)target8;
+       }
+       else if ((sparc_inst_op (ins) == 2) && (sparc_inst_op3 (ins) == 0x7)) {
+               /* sparc_set case 5. */
+               guint32 *p = c;
+
+               g_assert (v64);
+               reg = sparc_inst_rd (c [3]);
+               sparc_set (p, target, reg);
+               while (p < (c + 6))
+                       sparc_nop (p);
        }
        else
                NOT_IMPLEMENTED;
@@ -2220,13 +1965,13 @@ guint32*
 mono_sparc_emit_save_lmf (guint32 *code, guint32 lmf_offset)
 {
        /* Save lmf_addr */
-       sparc_st_imm (code, sparc_o0, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr));
+       sparc_sti_imm (code, sparc_o0, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr));
        /* Save previous_lmf */
-       sparc_ld (code, sparc_o0, sparc_g0, sparc_o7);
-       sparc_st_imm (code, sparc_o7, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf));
+       sparc_ldi (code, sparc_o0, sparc_g0, sparc_o7);
+       sparc_sti_imm (code, sparc_o7, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf));
        /* Set new lmf */
        sparc_add_imm (code, FALSE, sparc_fp, lmf_offset, sparc_o7);
-       sparc_st (code, sparc_o7, sparc_o0, sparc_g0);
+       sparc_sti (code, sparc_o7, sparc_o0, sparc_g0);
 
        return code;
 }
@@ -2235,11 +1980,11 @@ guint32*
 mono_sparc_emit_restore_lmf (guint32 *code, guint32 lmf_offset)
 {
        /* Load previous_lmf */
-       sparc_ld_imm (code, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sparc_l0);
+       sparc_ldi_imm (code, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sparc_l0);
        /* Load lmf_addr */
-       sparc_ld_imm (code, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sparc_l1);
+       sparc_ldi_imm (code, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sparc_l1);
        /* *(lmf) = previous_lmf */
-       sparc_st (code, sparc_l0, sparc_l1, sparc_g0);
+       sparc_sti (code, sparc_l0, sparc_l1, sparc_g0);
        return code;
 }
 
@@ -2251,17 +1996,17 @@ emit_save_sp_to_lmf (MonoCompile *cfg, guint32 *code)
         * set the sp field in the lmf before the call, not in the prolog.
         */
        if (cfg->method->save_lmf) {
-               gint32 lmf_offset = - cfg->arch.lmf_offset;
+               gint32 lmf_offset = MONO_SPARC_STACK_BIAS - cfg->arch.lmf_offset;
 
                /* Save sp */
-               sparc_st_imm (code, sparc_sp, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, sp));
+               sparc_sti_imm (code, sparc_sp, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, sp));
        }
 
        return code;
 }
 
 static guint32*
-emit_vret_token (MonoInst *ins, guint32 *code)
+emit_vret_token (MonoGenericSharingContext *gsctx, MonoInst *ins, guint32 *code)
 {
        MonoCallInst *call = (MonoCallInst*)ins;
        guint32 size;
@@ -2272,7 +2017,7 @@ emit_vret_token (MonoInst *ins, guint32 *code)
         */
        if (call->signature->pinvoke && MONO_TYPE_ISSTRUCT(call->signature->ret)) {
                if (call->signature->ret->type == MONO_TYPE_TYPEDBYREF)
-                       size = mono_type_stack_size (call->signature->ret, NULL);
+                       size = mini_type_stack_size (gsctx, call->signature->ret, NULL);
                else
                        size = mono_class_native_size (call->signature->ret->data.klass, NULL);
                sparc_unimp (code, size & 0xfff);
@@ -2285,39 +2030,54 @@ static guint32*
 emit_move_return_value (MonoInst *ins, guint32 *code)
 {
        /* Move return value to the target register */
-       /* FIXME: do this in the local reg allocator */
+       /* FIXME: do more things in the local reg allocator */
        switch (ins->opcode) {
        case OP_VOIDCALL:
        case OP_VOIDCALL_REG:
        case OP_VOIDCALL_MEMBASE:
                break;
-       case CEE_CALL:
+       case OP_CALL:
        case OP_CALL_REG:
        case OP_CALL_MEMBASE:
-               sparc_mov_reg_reg (code, sparc_o0, ins->dreg);
+               g_assert (ins->dreg == sparc_o0);
                break;
        case OP_LCALL:
        case OP_LCALL_REG:
        case OP_LCALL_MEMBASE:
                /* 
                 * ins->dreg is the least significant reg due to the lreg: LCALL rule
-                * in inssel.brg.
+                * in inssel-long32.brg.
                 */
-               sparc_mov_reg_reg (code, sparc_o0, ins->dreg + 1);
-               sparc_mov_reg_reg (code, sparc_o1, ins->dreg);
+#ifdef SPARCV9
+               sparc_mov_reg_reg (code, sparc_o0, ins->dreg);
+#else
+               g_assert (ins->dreg == sparc_o1);
+#endif
                break;
        case OP_FCALL:
        case OP_FCALL_REG:
        case OP_FCALL_MEMBASE:
+#ifdef SPARCV9
+               if (((MonoCallInst*)ins)->signature->ret->type == MONO_TYPE_R4) {
+                       sparc_fmovs (code, sparc_f0, ins->dreg);
+                       sparc_fstod (code, ins->dreg, ins->dreg);
+               }
+               else
+                       sparc_fmovd (code, sparc_f0, ins->dreg);
+#else          
                sparc_fmovs (code, sparc_f0, ins->dreg);
                if (((MonoCallInst*)ins)->signature->ret->type == MONO_TYPE_R4)
                        sparc_fstod (code, ins->dreg, ins->dreg);
                else
                        sparc_fmovs (code, sparc_f1, ins->dreg + 1);
+#endif
                break;
        case OP_VCALL:
        case OP_VCALL_REG:
        case OP_VCALL_MEMBASE:
+       case OP_VCALL2:
+       case OP_VCALL2_REG:
+       case OP_VCALL2_MEMBASE:
                break;
        default:
                NOT_IMPLEMENTED;
@@ -2343,34 +2103,36 @@ emit_load_volatile_arguments (MonoCompile *cfg, guint32 *code)
 
        /* FIXME: Generate intermediate code instead */
 
-       sig = method->signature;
+       sig = mono_method_signature (method);
 
-       cinfo = get_call_info (sig, FALSE);
+       cinfo = get_call_info (cfg, sig, FALSE);
        
        /* This is the opposite of the code in emit_prolog */
 
        for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
                ArgInfo *ainfo = cinfo->args + i;
-               guint32 stack_offset;
+               gint32 stack_offset;
                MonoType *arg_type;
-               inst = cfg->varinfo [i];
+
+               inst = cfg->args [i];
 
                if (sig->hasthis && (i == 0))
-                       arg_type = mono_defaults.object_class;
+                       arg_type = &mono_defaults.object_class->byval_arg;
                else
                        arg_type = sig->params [i - sig->hasthis];
 
-               stack_offset = ainfo->offset + 68;
+               stack_offset = ainfo->offset + ARGS_OFFSET;
                ireg = sparc_i0 + ainfo->reg;
 
                if (ainfo->storage == ArgInSplitRegStack) {
                        g_assert (inst->opcode == OP_REGOFFSET);
+
                        if (!sparc_is_imm13 (stack_offset))
                                NOT_IMPLEMENTED;
                        sparc_st_imm (code, inst->inst_basereg, stack_offset, sparc_i5);
                }
 
-               if (!arg_type->byref && (arg_type->type == MONO_TYPE_R8)) {
+               if (!v64 && !arg_type->byref && (arg_type->type == MONO_TYPE_R8)) {
                        if (ainfo->storage == ArgInIRegPair) {
                                if (!sparc_is_imm13 (inst->inst_offset + 4))
                                        NOT_IMPLEMENTED;
@@ -2398,7 +2160,7 @@ emit_load_volatile_arguments (MonoCompile *cfg, guint32 *code)
 
                                        }
                                }
-                       else
+                        else
                                g_assert_not_reached ();
                }
                else
@@ -2406,23 +2168,35 @@ emit_load_volatile_arguments (MonoCompile *cfg, guint32 *code)
                                /* Argument in register, but need to be saved to stack */
                                if (!sparc_is_imm13 (stack_offset))
                                        NOT_IMPLEMENTED;
-                               if (stack_offset & 0x1)
+                               if ((stack_offset - ARGS_OFFSET) & 0x1)
                                        /* FIXME: Is this ldsb or ldub ? */
                                        sparc_ldsb_imm (code, inst->inst_basereg, stack_offset, ireg);
                                else
-                                       if (stack_offset & 0x2)
+                                       if ((stack_offset - ARGS_OFFSET) & 0x2)
                                                sparc_ldsh_imm (code, inst->inst_basereg, stack_offset, ireg);
                                else
-                                       sparc_ld_imm (code, inst->inst_basereg, stack_offset, ireg);
+                                       if ((stack_offset - ARGS_OFFSET) & 0x4)
+                                               sparc_ld_imm (code, inst->inst_basereg, stack_offset, ireg);
+                                       else {
+                                               if (v64)
+                                                       sparc_ldx_imm (code, inst->inst_basereg, stack_offset, ireg);
+                                               else
+                                                       sparc_ld_imm (code, inst->inst_basereg, stack_offset, ireg);
+                                       }
                        }
-               else
-                       if ((ainfo->storage == ArgInIRegPair) && (inst->opcode != OP_REGVAR)) {
+                       else if ((ainfo->storage == ArgInIRegPair) && (inst->opcode != OP_REGVAR)) {
                                /* Argument in regpair, but need to be saved to stack */
                                if (!sparc_is_imm13 (inst->inst_offset + 4))
                                        NOT_IMPLEMENTED;
                                sparc_ld_imm (code, inst->inst_basereg, inst->inst_offset, ireg);
                                sparc_st_imm (code, inst->inst_basereg, inst->inst_offset + 4, ireg + 1);
                        }
+                       else if ((ainfo->storage == ArgInFloatReg) && (inst->opcode != OP_REGVAR)) {
+                               NOT_IMPLEMENTED;
+                       }
+                       else if ((ainfo->storage == ArgInDoubleReg) && (inst->opcode != OP_REGVAR)) {
+                               NOT_IMPLEMENTED;
+                       }
 
                if ((ainfo->storage == ArgInSplitRegStack) || (ainfo->storage == ArgOnStack))
                        if (inst->opcode == OP_REGVAR)
@@ -2466,40 +2240,205 @@ mono_sparc_is_virtual_call (guint32 *code)
 }
 
 /*
- * mono_sparc_get_vcall_slot_addr:
+ * mono_arch_get_vcall_slot:
  *
  *  Determine the vtable slot used by a virtual call.
  */
-gpointer*
-mono_sparc_get_vcall_slot_addr (guint32 *code, guint32 *fp)
+gpointer
+mono_arch_get_vcall_slot (guint8 *code8, gpointer *regs, int *displacement)
 {
+       guint32 *code = (guint32*)(gpointer)code8;
        guint32 ins = code [0];
        guint32 prev_ins = code [-1];
 
        mono_sparc_flushw ();
 
+       *displacement = 0;
+
+       if (!mono_sparc_is_virtual_call (code))
+               return NULL;
+
        if ((sparc_inst_op (ins) == 0x2) && (sparc_inst_op3 (ins) == 0x38)) {
-               if ((sparc_inst_op (prev_ins) == 0x3) && (sparc_inst_op3 (prev_ins) == 0)) {
+               if ((sparc_inst_op (prev_ins) == 0x3) && (sparc_inst_i (prev_ins) == 1) && (sparc_inst_op3 (prev_ins) == 0 || sparc_inst_op3 (prev_ins) == 0xb)) {
                        /* ld [r1 + CONST ], r2; call r2 */
                        guint32 base = sparc_inst_rs1 (prev_ins);
-                       guint32 disp = sparc_inst_imm13 (prev_ins);
-                       guint32 base_val;
+                       gint32 disp = (((gint32)(sparc_inst_imm13 (prev_ins))) << 19) >> 19;
+                       gpointer base_val;
 
                        g_assert (sparc_inst_rd (prev_ins) == sparc_inst_rs1 (ins));
 
                        g_assert ((base >= sparc_o0) && (base <= sparc_i7));
-                       
-                       base_val = fp [base - 16];
 
-                       return (gpointer)((guint8*)base_val + disp);
+                       base_val = regs [base];
+
+                       *displacement = disp;
+
+                       return (gpointer)base_val;
                }
-               else
+               else if ((sparc_inst_op (prev_ins) == 0x3) && (sparc_inst_i (prev_ins) == 0) && (sparc_inst_op3 (prev_ins) == 0)) {
+                       /* set r1, ICONST; ld [r1 + r2], r2; call r2 */
+                       /* Decode a sparc_set32 */
+                       guint32 base = sparc_inst_rs1 (prev_ins);
+                       guint32 disp;
+                       gpointer base_val;
+                       guint32 s1 = code [-3];
+                       guint32 s2 = code [-2];
+
+#ifdef SPARCV9
+                       NOT_IMPLEMENTED;
+#endif
+
+                       /* sparc_sethi */
+                       g_assert (sparc_inst_op (s1) == 0);
+                       g_assert (sparc_inst_op2 (s1) == 4);
+
+                       /* sparc_or_imm */
+                       g_assert (sparc_inst_op (s2) == 2);
+                       g_assert (sparc_inst_op3 (s2) == 2);
+                       g_assert (sparc_inst_i (s2) == 1);
+                       g_assert (sparc_inst_rs1 (s2) == sparc_inst_rd (s2));
+                       g_assert (sparc_inst_rd (s1) == sparc_inst_rs1 (s2));
+
+                       disp = ((s1 & 0x3fffff) << 10) | sparc_inst_imm13 (s2);
+
+                       g_assert ((base >= sparc_o0) && (base <= sparc_i7));
+
+                       base_val = regs [base];
+
+                       *displacement = disp;
+
+                       return (gpointer)base_val;
+               } else
                        g_assert_not_reached ();
        }
        else
                g_assert_not_reached ();
 
-       return FALSE;
+       return NULL;
+}
+
+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);
+}
+
+#define CMP_SIZE 3
+#define BR_SMALL_SIZE 2
+#define BR_LARGE_SIZE 2
+#define JUMP_IMM_SIZE 5
+#define ENABLE_WRONG_METHOD_CHECK 0
+
+/*
+ * LOCKING: called with the domain lock held
+ */
+gpointer
+mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count,
+       gpointer fail_tramp)
+{
+       int i;
+       int size = 0;
+       guint32 *code, *start;
+
+       g_assert (!fail_tramp);
+
+       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_domain_code_reserve (domain, size * 4);
+       start = code;
+
+       for (i = 0; i < count; ++i) {
+               MonoIMTCheckItem *item = imt_entries [i];
+               item->code_target = (guint8*)code;
+               if (item->is_equals) {
+                       if (item->check_target_idx) {
+                               if (!item->compare_done) {
+                                       sparc_set (code, (guint32)item->key, sparc_g5);
+                                       sparc_cmp (code, MONO_ARCH_IMT_REG, sparc_g5);
+                               }
+                               item->jmp_code = (guint8*)code;
+                               sparc_branch (code, 0, sparc_bne, 0);
+                               sparc_nop (code);
+                               sparc_set (code, ((guint32)(&(vtable->vtable [item->value.vtable_slot]))), sparc_g5);
+                               sparc_ld (code, sparc_g5, 0, sparc_g5);
+                               sparc_jmpl (code, sparc_g5, sparc_g0, sparc_g0);
+                               sparc_nop (code);
+                       } else {
+                               /* enable the commented code to assert on wrong method */
+#if ENABLE_WRONG_METHOD_CHECK
+                               g_assert_not_reached ();
+#endif
+                               sparc_set (code, ((guint32)(&(vtable->vtable [item->value.vtable_slot]))), sparc_g5);
+                               sparc_ld (code, sparc_g5, 0, sparc_g5);
+                               sparc_jmpl (code, sparc_g5, sparc_g0, sparc_g0);
+                               sparc_nop (code);
+#if ENABLE_WRONG_METHOD_CHECK
+                               g_assert_not_reached ();
+#endif
+                       }
+               } else {
+                       sparc_set (code, (guint32)item->key, sparc_g5);
+                       sparc_cmp (code, MONO_ARCH_IMT_REG, sparc_g5);
+                       item->jmp_code = (guint8*)code;
+                       sparc_branch (code, 0, sparc_beu, 0);
+                       sparc_nop (code);
+               }
+       }
+       /* 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) {
+                               sparc_patch ((guint32*)item->jmp_code, imt_entries [item->check_target_idx]->code_target);
+                       }
+               }
+       }
+
+       mono_arch_flush_icache ((guint8*)start, (code - start) * 4);
+
+       mono_stats.imt_thunks_size += (code - start) * 4;
+       g_assert (code - start <= size);
+       return start;
+}
+
+MonoMethod*
+mono_arch_find_imt_method (gpointer *regs, guint8 *code)
+{
+#ifdef SPARCV9
+       g_assert_not_reached ();
+#endif
+
+       return (MonoMethod*)regs [sparc_g1];
+}
+
+MonoObject*
+mono_arch_find_this_argument (gpointer *regs, MonoMethod *method, MonoGenericSharingContext *gsctx)
+{
+       mono_sparc_flushw ();
+
+       return (gpointer)regs [sparc_o0];
 }
 
 /*
@@ -2517,9 +2456,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
        guint32 *code = (guint32*)(cfg->native_code + cfg->code_len);
        MonoInst *last_ins = NULL;
        int max_len, cpos;
-
-       if (cfg->opt & MONO_OPT_PEEPHOLE)
-               peephole_pass (cfg, bb);
+       const char *spec;
 
        if (cfg->verbose_level > 2)
                g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
@@ -2530,13 +2467,14 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                NOT_IMPLEMENTED;
        }
 
-       ins = bb->code;
-       while (ins) {
+       MONO_BB_FOR_EACH_INS (bb, ins) {
                guint8* code_start;
 
                offset = (guint8*)code - cfg->native_code;
 
-               max_len = ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
+               spec = ins_get_spec (ins->opcode);
+
+               max_len = ((guint8 *)spec)[MONO_INST_LEN];
 
                if (offset > (cfg->code_size - max_len - 16)) {
                        cfg->code_size *= 2;
@@ -2556,14 +2494,20 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        EMIT_STORE_MEMBASE_IMM (ins, sth);
                        break;
                case OP_STORE_MEMBASE_IMM:
+                       EMIT_STORE_MEMBASE_IMM (ins, sti);
+                       break;
                case OP_STOREI4_MEMBASE_IMM:
                        EMIT_STORE_MEMBASE_IMM (ins, st);
                        break;
                case OP_STOREI8_MEMBASE_IMM:
+#ifdef SPARCV9
+                       EMIT_STORE_MEMBASE_IMM (ins, stx);
+#else
                        /* Only generated by peephole opts */
                        g_assert ((ins->inst_offset % 8) == 0);
                        g_assert (ins->inst_imm == 0);
                        EMIT_STORE_MEMBASE_IMM (ins, stx);
+#endif
                        break;
                case OP_STOREI1_MEMBASE_REG:
                        EMIT_STORE_MEMBASE_REG (ins, stb);
@@ -2571,26 +2515,31 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_STOREI2_MEMBASE_REG:
                        EMIT_STORE_MEMBASE_REG (ins, sth);
                        break;
-               case OP_STORE_MEMBASE_REG:
                case OP_STOREI4_MEMBASE_REG:
                        EMIT_STORE_MEMBASE_REG (ins, st);
                        break;
                case OP_STOREI8_MEMBASE_REG:
+#ifdef SPARCV9
+                       EMIT_STORE_MEMBASE_REG (ins, stx);
+#else
                        /* Only used by OP_MEMSET */
                        EMIT_STORE_MEMBASE_REG (ins, std);
+#endif
                        break;
-               case CEE_LDIND_I:
-               case CEE_LDIND_I4:
-               case CEE_LDIND_U4:
-                       sparc_ld (code, ins->inst_p0, sparc_g0, ins->dreg);
+               case OP_STORE_MEMBASE_REG:
+                       EMIT_STORE_MEMBASE_REG (ins, sti);
                        break;
-               /* The cast IS BAD (maybe).  But it needs to be done... */
                case OP_LOADU4_MEM:
-                       sparc_set (code, (guint)ins->inst_p0, ins->dreg);
+                       sparc_set (code, ins->inst_c0, ins->dreg);
                        sparc_ld (code, ins->dreg, sparc_g0, ins->dreg);
                        break;
-               case OP_LOAD_MEMBASE:
                case OP_LOADI4_MEMBASE:
+#ifdef SPARCV9
+                       EMIT_LOAD_MEMBASE (ins, ldsw);
+#else
+                       EMIT_LOAD_MEMBASE (ins, ld);
+#endif
+                       break;
                case OP_LOADU4_MEMBASE:
                        EMIT_LOAD_MEMBASE (ins, ld);
                        break;
@@ -2606,26 +2555,62 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_LOADI2_MEMBASE:
                        EMIT_LOAD_MEMBASE (ins, ldsh);
                        break;
-               case CEE_CONV_I1:
+               case OP_LOAD_MEMBASE:
+#ifdef SPARCV9
+                               EMIT_LOAD_MEMBASE (ins, ldx);
+#else
+                               EMIT_LOAD_MEMBASE (ins, ld);
+#endif
+                       break;
+#ifdef SPARCV9
+               case OP_LOADI8_MEMBASE:
+                       EMIT_LOAD_MEMBASE (ins, ldx);
+                       break;
+#endif
+               case OP_ICONV_TO_I1:
                        sparc_sll_imm (code, ins->sreg1, 24, sparc_o7);
                        sparc_sra_imm (code, sparc_o7, 24, ins->dreg);
                        break;
-               case CEE_CONV_I2:
+               case OP_ICONV_TO_I2:
                        sparc_sll_imm (code, ins->sreg1, 16, sparc_o7);
                        sparc_sra_imm (code, sparc_o7, 16, ins->dreg);
                        break;
-               /* GCC does this one differently.  Don't ask me WHY. */
-               case CEE_CONV_U1:
+               case OP_ICONV_TO_U1:
                        sparc_and_imm (code, FALSE, ins->sreg1, 0xff, ins->dreg);
                        break;
-               case CEE_CONV_U2:
+               case OP_ICONV_TO_U2:
                        sparc_sll_imm (code, ins->sreg1, 16, sparc_o7);
                        sparc_srl_imm (code, sparc_o7, 16, ins->dreg);
                        break;
+               case OP_LCONV_TO_OVF_U4:
+               case OP_ICONV_TO_OVF_U4:
+                       /* Only used on V9 */
+                       sparc_cmp_imm (code, ins->sreg1, 0);
+                       mono_add_patch_info (cfg, (guint8*)(code) - (cfg)->native_code,
+                                                                MONO_PATCH_INFO_EXC, "OverflowException");
+                       sparc_branchp (code, 0, sparc_bl, sparc_xcc_short, 0, 0);
+                       /* Delay slot */
+                       sparc_set (code, 1, sparc_o7);
+                       sparc_sllx_imm (code, sparc_o7, 32, sparc_o7);
+                       sparc_cmp (code, ins->sreg1, sparc_o7);
+                       mono_add_patch_info (cfg, (guint8*)(code) - (cfg)->native_code,
+                                                                MONO_PATCH_INFO_EXC, "OverflowException");
+                       sparc_branchp (code, 0, sparc_bge, sparc_xcc_short, 0, 0);
+                       sparc_nop (code);
+                       sparc_mov_reg_reg (code, ins->sreg1, ins->dreg);
+                       break;
+               case OP_LCONV_TO_OVF_I4_UN:
+               case OP_ICONV_TO_OVF_I4_UN:
+                       /* Only used on V9 */
+                       NOT_IMPLEMENTED;
+                       break;
                case OP_COMPARE:
+               case OP_LCOMPARE:
+               case OP_ICOMPARE:
                        sparc_cmp (code, ins->sreg1, ins->sreg2);
                        break;
                case OP_COMPARE_IMM:
+               case OP_ICOMPARE_IMM:
                        if (sparc_is_imm13 (ins->inst_imm))
                                sparc_cmp_imm (code, ins->sreg1, ins->inst_imm);
                        else {
@@ -2633,73 +2618,80 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                sparc_cmp (code, ins->sreg1, sparc_o7);
                        }
                        break;
-               case OP_X86_TEST_NULL:
-                       sparc_cmp_imm (code, ins->sreg1, 0);
-                       break;
-               case CEE_BREAK:
+               case OP_BREAK:
                        /*
                         * gdb does not like encountering 'ta 1' in the debugged code. So 
                         * instead of emitting a trap, we emit a call a C function and place a 
                         * breakpoint there.
                         */
-                       mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_ABS, mono_sparc_break);
-                       sparc_call_simple (code, 0);
-                       sparc_nop (code);
+                       //sparc_ta (code, 1);
+                       mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_ABS, mono_break);
+                       EMIT_CALL();
                        break;
                case OP_ADDCC:
+               case OP_IADDCC:
                        sparc_add (code, TRUE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
-               case CEE_ADD:
+               case OP_IADD:
                        sparc_add (code, FALSE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_ADDCC_IMM:
                case OP_ADD_IMM:
+               case OP_IADD_IMM:
                        /* according to inssel-long32.brg, this should set cc */
                        EMIT_ALU_IMM (ins, add, TRUE);
                        break;
                case OP_ADC:
+               case OP_IADC:
                        /* according to inssel-long32.brg, this should set cc */
                        sparc_addx (code, TRUE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_ADC_IMM:
+               case OP_IADC_IMM:
                        EMIT_ALU_IMM (ins, addx, TRUE);
                        break;
                case OP_SUBCC:
+               case OP_ISUBCC:
                        sparc_sub (code, TRUE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
-               case CEE_SUB:
+               case OP_ISUB:
                        sparc_sub (code, FALSE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_SUBCC_IMM:
                case OP_SUB_IMM:
+               case OP_ISUB_IMM:
                        /* according to inssel-long32.brg, this should set cc */
                        EMIT_ALU_IMM (ins, sub, TRUE);
                        break;
                case OP_SBB:
+               case OP_ISBB:
                        /* according to inssel-long32.brg, this should set cc */
                        sparc_subx (code, TRUE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_SBB_IMM:
+               case OP_ISBB_IMM:
                        EMIT_ALU_IMM (ins, subx, TRUE);
                        break;
-               case CEE_AND:
+               case OP_IAND:
                        sparc_and (code, FALSE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_AND_IMM:
+               case OP_IAND_IMM:
                        EMIT_ALU_IMM (ins, and, FALSE);
                        break;
-               case CEE_DIV:
+               case OP_IDIV:
                        /* Sign extend sreg1 into %y */
                        sparc_sra_imm (code, ins->sreg1, 31, sparc_o7);
                        sparc_wry (code, sparc_o7, sparc_g0);
                        sparc_sdiv (code, TRUE, ins->sreg1, ins->sreg2, ins->dreg);
-                       EMIT_COND_SYSTEM_EXCEPTION (code, sparc_boverflow, "ArithmeticException");
+                       EMIT_COND_SYSTEM_EXCEPTION_GENERAL (code, sparc_boverflow, "ArithmeticException", TRUE, sparc_icc_short);
                        break;
-               case CEE_DIV_UN:
+               case OP_IDIV_UN:
                        sparc_wry (code, sparc_g0, sparc_g0);
                        sparc_udiv (code, FALSE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
-               case OP_DIV_IMM: {
+               case OP_DIV_IMM:
+               case OP_IDIV_IMM: {
                        int i, imm;
 
                        /* Transform division into a shift */
@@ -2728,70 +2720,79 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                sparc_sra_imm (code, ins->sreg1, 31, sparc_o7);
                                sparc_wry (code, sparc_o7, sparc_g0);
                                EMIT_ALU_IMM (ins, sdiv, TRUE);
-                               EMIT_COND_SYSTEM_EXCEPTION (code, sparc_boverflow, "ArithmeticException");
+                               EMIT_COND_SYSTEM_EXCEPTION_GENERAL (code, sparc_boverflow, "ArithmeticException", TRUE, sparc_icc_short);
                        }
                        break;
                }
-               case CEE_REM:
+               case OP_IDIV_UN_IMM:
+                       sparc_wry (code, sparc_g0, sparc_g0);
+                       EMIT_ALU_IMM (ins, udiv, FALSE);
+                       break;
+               case OP_IREM:
                        /* Sign extend sreg1 into %y */
                        sparc_sra_imm (code, ins->sreg1, 31, sparc_o7);
                        sparc_wry (code, sparc_o7, sparc_g0);
                        sparc_sdiv (code, TRUE, ins->sreg1, ins->sreg2, sparc_o7);
-                       EMIT_COND_SYSTEM_EXCEPTION (code, sparc_boverflow, "ArithmeticException");
+                       EMIT_COND_SYSTEM_EXCEPTION_GENERAL (code, sparc_boverflow, "ArithmeticException", TRUE, sparc_icc_short);
                        sparc_smul (code, FALSE, ins->sreg2, sparc_o7, sparc_o7);
                        sparc_sub (code, FALSE, ins->sreg1, sparc_o7, ins->dreg);
                        break;
-               case CEE_REM_UN:
+               case OP_IREM_UN:
                        sparc_wry (code, sparc_g0, sparc_g0);
                        sparc_udiv (code, FALSE, ins->sreg1, ins->sreg2, sparc_o7);
                        sparc_umul (code, FALSE, ins->sreg2, sparc_o7, sparc_o7);
                        sparc_sub (code, FALSE, ins->sreg1, sparc_o7, ins->dreg);
                        break;
                case OP_REM_IMM:
+               case OP_IREM_IMM:
                        /* Sign extend sreg1 into %y */
                        sparc_sra_imm (code, ins->sreg1, 31, sparc_o7);
                        sparc_wry (code, sparc_o7, sparc_g0);
                        if (!sparc_is_imm13 (ins->inst_imm)) {
-                               sparc_set (code, ins->inst_imm, sparc_g1);
-                               sparc_sdiv (code, TRUE, ins->sreg1, sparc_g1, sparc_o7);
-                               EMIT_COND_SYSTEM_EXCEPTION (code, sparc_boverflow, "ArithmeticException");
-                               sparc_smul (code, FALSE, sparc_o7, sparc_g1, sparc_o7);
+                               sparc_set (code, ins->inst_imm, GP_SCRATCH_REG);
+                               sparc_sdiv (code, TRUE, ins->sreg1, GP_SCRATCH_REG, sparc_o7);
+                               EMIT_COND_SYSTEM_EXCEPTION_GENERAL (code, sparc_boverflow, "ArithmeticException", TRUE, sparc_icc_short);
+                               sparc_smul (code, FALSE, sparc_o7, GP_SCRATCH_REG, sparc_o7);
                        }
                        else {
                                sparc_sdiv_imm (code, TRUE, ins->sreg1, ins->inst_imm, sparc_o7);
-                               EMIT_COND_SYSTEM_EXCEPTION (code, sparc_boverflow, "ArithmeticException");
+                               EMIT_COND_SYSTEM_EXCEPTION_GENERAL (code, sparc_boverflow, "ArithmeticException", TRUE, sparc_icc_short);
                                sparc_smul_imm (code, FALSE, sparc_o7, ins->inst_imm, sparc_o7);
                        }
                        sparc_sub (code, FALSE, ins->sreg1, sparc_o7, ins->dreg);
                        break;
-               case CEE_OR:
+               case OP_IOR:
                        sparc_or (code, FALSE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_OR_IMM:
+               case OP_IOR_IMM:
                        EMIT_ALU_IMM (ins, or, FALSE);
                        break;
-               case CEE_XOR:
+               case OP_IXOR:
                        sparc_xor (code, FALSE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_XOR_IMM:
+               case OP_IXOR_IMM:
                        EMIT_ALU_IMM (ins, xor, FALSE);
                        break;
-               case CEE_SHL:
+               case OP_ISHL:
                        sparc_sll (code, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
                case OP_SHL_IMM:
-                       if (sparc_is_imm13 (ins->inst_imm))
+               case OP_ISHL_IMM:
+                       if (ins->inst_imm < (1 << 5))
                                sparc_sll_imm (code, ins->sreg1, ins->inst_imm, ins->dreg);
                        else {
                                sparc_set (code, ins->inst_imm, sparc_o7);
                                sparc_sll (code, ins->sreg1, sparc_o7, ins->dreg);
                        }
                        break;
-               case CEE_SHR:
+               case OP_ISHR:
                        sparc_sra (code, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
+               case OP_ISHR_IMM:
                case OP_SHR_IMM:
-                       if (sparc_is_imm13 (ins->inst_imm))
+                       if (ins->inst_imm < (1 << 5))
                                sparc_sra_imm (code, ins->sreg1, ins->inst_imm, ins->dreg);
                        else {
                                sparc_set (code, ins->inst_imm, sparc_o7);
@@ -2799,32 +2800,70 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        }
                        break;
                case OP_SHR_UN_IMM:
-                       if (sparc_is_imm13 (ins->inst_imm))
+               case OP_ISHR_UN_IMM:
+                       if (ins->inst_imm < (1 << 5))
                                sparc_srl_imm (code, ins->sreg1, ins->inst_imm, ins->dreg);
                        else {
                                sparc_set (code, ins->inst_imm, sparc_o7);
                                sparc_srl (code, ins->sreg1, sparc_o7, ins->dreg);
                        }
                        break;
-               case CEE_SHR_UN:
+               case OP_ISHR_UN:
                        sparc_srl (code, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
-               case CEE_NOT:
+               case OP_LSHL:
+                       sparc_sllx (code, ins->sreg1, ins->sreg2, ins->dreg);
+                       break;
+               case OP_LSHL_IMM:
+                       if (ins->inst_imm < (1 << 6))
+                               sparc_sllx_imm (code, ins->sreg1, ins->inst_imm, ins->dreg);
+                       else {
+                               sparc_set (code, ins->inst_imm, sparc_o7);
+                               sparc_sllx (code, ins->sreg1, sparc_o7, ins->dreg);
+                       }
+                       break;
+               case OP_LSHR:
+                       sparc_srax (code, ins->sreg1, ins->sreg2, ins->dreg);
+                       break;
+               case OP_LSHR_IMM:
+                       if (ins->inst_imm < (1 << 6))
+                               sparc_srax_imm (code, ins->sreg1, ins->inst_imm, ins->dreg);
+                       else {
+                               sparc_set (code, ins->inst_imm, sparc_o7);
+                               sparc_srax (code, ins->sreg1, sparc_o7, ins->dreg);
+                       }
+                       break;
+               case OP_LSHR_UN:
+                       sparc_srlx (code, ins->sreg1, ins->sreg2, ins->dreg);
+                       break;
+               case OP_LSHR_UN_IMM:
+                       if (ins->inst_imm < (1 << 6))
+                               sparc_srlx_imm (code, ins->sreg1, ins->inst_imm, ins->dreg);
+                       else {
+                               sparc_set (code, ins->inst_imm, sparc_o7);
+                               sparc_srlx (code, ins->sreg1, sparc_o7, ins->dreg);
+                       }
+                       break;
+               case OP_INOT:
                        /* can't use sparc_not */
                        sparc_xnor (code, FALSE, ins->sreg1, sparc_g0, ins->dreg);
                        break;
-               case CEE_NEG:
+               case OP_INEG:
                        /* can't use sparc_neg */
                        sparc_sub (code, FALSE, sparc_g0, ins->sreg1, ins->dreg);
                        break;
-               case CEE_MUL:
+               case OP_IMUL:
                        sparc_smul (code, FALSE, ins->sreg1, ins->sreg2, ins->dreg);
                        break;
+               case OP_IMUL_IMM:
                case OP_MUL_IMM: {
                        int i, imm;
 
+                       if ((ins->inst_imm == 1) && (ins->sreg1 == ins->dreg))
+                               break;
+
                        /* Transform multiplication into a shift */
-                       for (i = 1; i < 30; ++i) {
+                       for (i = 0; i < 30; ++i) {
                                imm = (1 << i);
                                if (ins->inst_imm == imm)
                                        break;
@@ -2835,74 +2874,89 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                EMIT_ALU_IMM (ins, smul, FALSE);
                        break;
                }
-               case CEE_MUL_OVF:
+               case OP_IMUL_OVF:
                        sparc_smul (code, TRUE, ins->sreg1, ins->sreg2, ins->dreg);
                        sparc_rdy (code, sparc_g1);
                        sparc_sra_imm (code, ins->dreg, 31, sparc_o7);
                        sparc_cmp (code, sparc_g1, sparc_o7);
-                       EMIT_COND_SYSTEM_EXCEPTION (ins, sparc_bne, "OverflowException");
+                       EMIT_COND_SYSTEM_EXCEPTION_GENERAL (ins, sparc_bne, "OverflowException", TRUE, sparc_icc_short);
                        break;
-               case CEE_MUL_OVF_UN:
+               case OP_IMUL_OVF_UN:
                        sparc_umul (code, TRUE, ins->sreg1, ins->sreg2, ins->dreg);
                        sparc_rdy (code, sparc_o7);
                        sparc_cmp (code, sparc_o7, sparc_g0);
-                       EMIT_COND_SYSTEM_EXCEPTION (ins, sparc_bne, "OverflowException");
+                       EMIT_COND_SYSTEM_EXCEPTION_GENERAL (ins, sparc_bne, "OverflowException", TRUE, sparc_icc_short);
                        break;
                case OP_ICONST:
-               case OP_SETREGIMM:
                        sparc_set (code, ins->inst_c0, ins->dreg);
                        break;
+               case OP_I8CONST:
+                       sparc_set (code, ins->inst_l, ins->dreg);
+                       break;
                case OP_AOTCONST:
                        mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
-                       sparc_set (code, 0xffffff, ins->dreg);
+                       sparc_set_template (code, ins->dreg);
+                       break;
+               case OP_JUMP_TABLE:
+                       mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
+                       sparc_set_template (code, ins->dreg);
                        break;
-               case CEE_CONV_I4:
-               case CEE_CONV_U4:
+               case OP_ICONV_TO_I4:
+               case OP_ICONV_TO_U4:
                case OP_MOVE:
-               case OP_SETREG:
                        if (ins->sreg1 != ins->dreg)
                                sparc_mov_reg_reg (code, ins->sreg1, ins->dreg);
                        break;
-               case CEE_JMP:
+               case OP_FMOVE:
+#ifdef SPARCV9
+                       if (ins->sreg1 != ins->dreg)
+                               sparc_fmovd (code, ins->sreg1, ins->dreg);
+#else
+                       sparc_fmovs (code, ins->sreg1, ins->dreg);
+                       sparc_fmovs (code, ins->sreg1 + 1, ins->dreg + 1);
+#endif
+                       break;
+               case OP_JMP:
                        if (cfg->method->save_lmf)
                                NOT_IMPLEMENTED;
 
                        code = emit_load_volatile_arguments (cfg, code);
                        mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
-                       sparc_set (code, 0xffffff, sparc_o7);
+                       sparc_set_template (code, sparc_o7);
                        sparc_jmpl (code, sparc_o7, sparc_g0, sparc_g0);
                        /* Restore parent frame in delay slot */
                        sparc_restore_imm (code, sparc_g0, 0, sparc_g0);
                        break;
                case OP_CHECK_THIS:
                        /* ensure ins->sreg1 is not NULL */
-                       sparc_ld_imm (code, ins->sreg1, 0, sparc_g0);
+                       /* Might be misaligned in case of vtypes so use a byte load */
+                       sparc_ldsb_imm (code, ins->sreg1, 0, sparc_g0);
                        break;
                case OP_ARGLIST:
                        sparc_add_imm (code, FALSE, sparc_fp, cfg->sig_cookie, sparc_o7);
-                       sparc_st_imm (code, sparc_o7, ins->sreg1, 0);
+                       sparc_sti_imm (code, sparc_o7, ins->sreg1, 0);
                        break;
                case OP_FCALL:
                case OP_LCALL:
                case OP_VCALL:
+               case OP_VCALL2:
                case OP_VOIDCALL:
-               case CEE_CALL:
+               case OP_CALL:
                        call = (MonoCallInst*)ins;
                        g_assert (!call->virtual);
                        code = emit_save_sp_to_lmf (cfg, code);
                        if (ins->flags & MONO_INST_HAS_METHOD)
-                               mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_METHOD, call->method);
+                           code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
                        else
-                               mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_ABS, call->fptr);
-                       sparc_call_simple (code, 0);
-                       sparc_nop (code);
+                           code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr);
 
-                       code = emit_vret_token (ins, code);
+                       code = emit_vret_token (cfg->generic_sharing_context, ins, code);
                        code = emit_move_return_value (ins, code);
                        break;
                case OP_FCALL_REG:
                case OP_LCALL_REG:
                case OP_VCALL_REG:
+               case OP_VCALL2_REG:
                case OP_VOIDCALL_REG:
                case OP_CALL_REG:
                        call = (MonoCallInst*)ins;
@@ -2918,51 +2972,115 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        else
                                sparc_nop (code);
 
-                       code = emit_vret_token (ins, code);
+                       code = emit_vret_token (cfg->generic_sharing_context, ins, code);
                        code = emit_move_return_value (ins, code);
                        break;
                case OP_FCALL_MEMBASE:
                case OP_LCALL_MEMBASE:
                case OP_VCALL_MEMBASE:
+               case OP_VCALL2_MEMBASE:
                case OP_VOIDCALL_MEMBASE:
                case OP_CALL_MEMBASE:
                        call = (MonoCallInst*)ins;
-                       g_assert (sparc_is_imm13 (ins->inst_offset));
                        code = emit_save_sp_to_lmf (cfg, code);
-                       sparc_ld_imm (code, ins->inst_basereg, ins->inst_offset, sparc_o7);
+                       if (sparc_is_imm13 (ins->inst_offset)) {
+                               sparc_ldi_imm (code, ins->inst_basereg, ins->inst_offset, sparc_o7);
+                       } else {
+                               sparc_set (code, ins->inst_offset, sparc_o7);
+                               sparc_ldi (code, ins->inst_basereg, sparc_o7, sparc_o7);
+                       }
                        sparc_jmpl (code, sparc_o7, sparc_g0, sparc_callsite);
                        if (call->virtual)
                                sparc_or_imm (code, FALSE, sparc_g0, 0xca, sparc_g0);
                        else
                                sparc_nop (code);
 
-                       code = emit_vret_token (ins, code);
+                       code = emit_vret_token (cfg->generic_sharing_context, ins, code);
                        code = emit_move_return_value (ins, code);
                        break;
                case OP_SETFRET:
-                       if (cfg->method->signature->ret->type == MONO_TYPE_R4)
+                       if (mono_method_signature (cfg->method)->ret->type == MONO_TYPE_R4)
                                sparc_fdtos (code, ins->sreg1, sparc_f0);
                        else {
+#ifdef SPARCV9
+                               sparc_fmovd (code, ins->sreg1, ins->dreg);
+#else
+                               /* FIXME: Why not use fmovd ? */
                                sparc_fmovs (code, ins->sreg1, ins->dreg);
                                sparc_fmovs (code, ins->sreg1 + 1, ins->dreg + 1);
+#endif
                        }
                        break;
-               case OP_OUTARG:
-                       g_assert_not_reached ();
-                       break;
-               case OP_LOCALLOC:
+               case OP_LOCALLOC: {
+                       guint32 size_reg;
+                       gint32 offset2;
+
+#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
+                       /* Perform stack touching */
+                       NOT_IMPLEMENTED;
+#endif
+
                        /* Keep alignment */
-                       sparc_add_imm (code, FALSE, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1, ins->dreg);
-                       sparc_set (code, ~(MONO_ARCH_FRAME_ALIGNMENT - 1), sparc_o7);
+                       /* Add 4 to compensate for the rounding of localloc_offset */
+                       sparc_add_imm (code, FALSE, ins->sreg1, 4 + MONO_ARCH_LOCALLOC_ALIGNMENT - 1, ins->dreg);
+                       sparc_set (code, ~(MONO_ARCH_LOCALLOC_ALIGNMENT - 1), sparc_o7);
                        sparc_and (code, FALSE, ins->dreg, sparc_o7, ins->dreg);
+
+                       if ((ins->flags & MONO_INST_INIT) && (ins->sreg1 == ins->dreg)) {
+#ifdef SPARCV9
+                               size_reg = sparc_g4;
+#else
+                               size_reg = sparc_g1;
+#endif
+                               sparc_mov_reg_reg (code, ins->dreg, size_reg);
+                       }
+                       else
+                               size_reg = ins->sreg1;
+
                        sparc_sub (code, FALSE, sparc_sp, ins->dreg, ins->dreg);
                        /* Keep %sp valid at all times */
                        sparc_mov_reg_reg (code, ins->dreg, sparc_sp);
-                       g_assert (sparc_is_imm13 (cfg->arch.localloc_offset));
-                       sparc_add_imm (code, FALSE, ins->dreg, cfg->arch.localloc_offset, ins->dreg);
+                       /* Round localloc_offset too so the result is at least 8 aligned */
+                       offset2 = ALIGN_TO (cfg->arch.localloc_offset, 8);
+                       g_assert (sparc_is_imm13 (MONO_SPARC_STACK_BIAS + offset2));
+                       sparc_add_imm (code, FALSE, ins->dreg, MONO_SPARC_STACK_BIAS + offset2, ins->dreg);
+
+                       if (ins->flags & MONO_INST_INIT) {
+                               guint32 *br [3];
+                               /* Initialize memory region */
+                               sparc_cmp_imm (code, size_reg, 0);
+                               br [0] = code;
+                               sparc_branch (code, 0, sparc_be, 0);
+                               /* delay slot */
+                               sparc_set (code, 0, sparc_o7);
+                               sparc_sub_imm (code, 0, size_reg, sparcv9 ? 8 : 4, size_reg);
+                               /* start of loop */
+                               br [1] = code;
+                               if (sparcv9)
+                                       sparc_stx (code, sparc_g0, ins->dreg, sparc_o7);
+                               else
+                                       sparc_st (code, sparc_g0, ins->dreg, sparc_o7);
+                               sparc_cmp (code, sparc_o7, size_reg);
+                               br [2] = code;
+                               sparc_branch (code, 0, sparc_bl, 0);
+                               sparc_patch (br [2], br [1]);
+                               /* delay slot */
+                               sparc_add_imm (code, 0, sparc_o7, sparcv9 ? 8 : 4, sparc_o7);
+                               sparc_patch (br [0], code);
+                       }
                        break;
-               case OP_SPARC_LOCALLOC_IMM: {
-                       guint32 offset = ins->inst_c0;
+               }
+               case OP_LOCALLOC_IMM: {
+                       gint32 offset = ins->inst_imm;
+                       gint32 offset2;
+                       
+#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
+                       /* Perform stack touching */
+                       NOT_IMPLEMENTED;
+#endif
+
+                       /* To compensate for the rounding of localloc_offset */
+                       offset += sizeof (gpointer);
                        offset = ALIGN_TO (offset, MONO_ARCH_FRAME_ALIGNMENT);
                        if (sparc_is_imm13 (offset))
                                sparc_sub_imm (code, FALSE, sparc_sp, offset, sparc_sp);
@@ -2970,21 +3088,57 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                sparc_set (code, offset, sparc_o7);
                                sparc_sub (code, FALSE, sparc_sp, sparc_o7, sparc_sp);
                        }
-                       sparc_mov_reg_reg (code, sparc_sp, ins->dreg);
-                       g_assert (sparc_is_imm13 (cfg->arch.localloc_offset));
-                       sparc_add_imm (code, FALSE, ins->dreg, cfg->arch.localloc_offset, ins->dreg);
+                       /* Round localloc_offset too so the result is at least 8 aligned */
+                       offset2 = ALIGN_TO (cfg->arch.localloc_offset, 8);
+                       g_assert (sparc_is_imm13 (MONO_SPARC_STACK_BIAS + offset2));
+                       sparc_add_imm (code, FALSE, sparc_sp, MONO_SPARC_STACK_BIAS + offset2, ins->dreg);
+                       if ((ins->flags & MONO_INST_INIT) && (offset > 0)) {
+                               guint32 *br [2];
+                               int i;
+
+                               if (offset <= 16) {
+                                       i = 0;
+                                       while (i < offset) {
+                                               if (sparcv9) {
+                                                       sparc_stx_imm (code, sparc_g0, ins->dreg, i);
+                                                       i += 8;
+                                               }
+                                               else {
+                                                       sparc_st_imm (code, sparc_g0, ins->dreg, i);
+                                                       i += 4;
+                                               }
+                                       }
+                               }
+                               else {
+                                       sparc_set (code, offset, sparc_o7);
+                                       sparc_sub_imm (code, 0, sparc_o7, sparcv9 ? 8 : 4, sparc_o7);
+                                       /* beginning of loop */
+                                       br [0] = code;
+                                       if (sparcv9)
+                                               sparc_stx (code, sparc_g0, ins->dreg, sparc_o7);
+                                       else
+                                               sparc_st (code, sparc_g0, ins->dreg, sparc_o7);
+                                       sparc_cmp_imm (code, sparc_o7, 0);
+                                       br [1] = code;
+                                       sparc_branch (code, 0, sparc_bne, 0);
+                                       /* delay slot */
+                                       sparc_sub_imm (code, 0, sparc_o7, sparcv9 ? 8 : 4, sparc_o7);
+                                       sparc_patch (br [1], br [0]);
+                               }
+                       }
                        break;
                }
-               case CEE_RET:
-                       /* The return is done in the epilog */
-                       g_assert_not_reached ();
-                       break;
-               case CEE_THROW:
-                       mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_INTERNAL_METHOD, 
+               case OP_THROW:
+                       sparc_mov_reg_reg (code, ins->sreg1, sparc_o0);
+                       mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, 
                                             (gpointer)"mono_arch_throw_exception");
-                       sparc_call_simple (code, 0);
-                       /* Delay slot */
+                       EMIT_CALL ();
+                       break;
+               case OP_RETHROW:
                        sparc_mov_reg_reg (code, ins->sreg1, sparc_o0);
+                       mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, 
+                                            (gpointer)"mono_arch_rethrow_exception");
+                       EMIT_CALL ();
                        break;
                case OP_START_HANDLER: {
                        /*
@@ -2997,47 +3151,55 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                         */
                        MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
                        if (!sparc_is_imm13 (spvar->inst_offset)) {
-                               sparc_set (code, spvar->inst_offset, sparc_g0);
-                               sparc_st (code, sparc_o7, spvar->inst_basereg, sparc_g0);
+                               sparc_set (code, spvar->inst_offset, GP_SCRATCH_REG);
+                               sparc_sti (code, sparc_o7, spvar->inst_basereg, GP_SCRATCH_REG);
                        }
                        else
-                               sparc_st_imm (code, sparc_o7, spvar->inst_basereg, spvar->inst_offset);
+                               sparc_sti_imm (code, sparc_o7, spvar->inst_basereg, spvar->inst_offset);
                        break;
                }
                case OP_ENDFILTER: {
                        MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
                        if (!sparc_is_imm13 (spvar->inst_offset)) {
-                               sparc_set (code, spvar->inst_offset, sparc_g0);
-                               sparc_ld (code, spvar->inst_basereg, sparc_g0, sparc_o7);
+                               sparc_set (code, spvar->inst_offset, GP_SCRATCH_REG);
+                               sparc_ldi (code, spvar->inst_basereg, GP_SCRATCH_REG, sparc_o7);
                        }
                        else
-                               sparc_ld_imm (code, spvar->inst_basereg, spvar->inst_offset, sparc_o7);
+                               sparc_ldi_imm (code, spvar->inst_basereg, spvar->inst_offset, sparc_o7);
                        sparc_jmpl_imm (code, sparc_o7, 8, sparc_g0);
                        /* Delay slot */
                        sparc_mov_reg_reg (code, ins->sreg1, sparc_o0);
                        break;
                }
-               case CEE_ENDFINALLY: {
+               case OP_ENDFINALLY: {
                        MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
                        if (!sparc_is_imm13 (spvar->inst_offset)) {
-                               sparc_set (code, spvar->inst_offset, sparc_g0);
-                               sparc_ld (code, spvar->inst_basereg, sparc_g0, sparc_o7);
+                               sparc_set (code, spvar->inst_offset, GP_SCRATCH_REG);
+                               sparc_ldi (code, spvar->inst_basereg, GP_SCRATCH_REG, sparc_o7);
                        }
                        else
-                               sparc_ld_imm (code, spvar->inst_basereg, spvar->inst_offset, sparc_o7);
+                               sparc_ldi_imm (code, spvar->inst_basereg, spvar->inst_offset, sparc_o7);
                        sparc_jmpl_imm (code, sparc_o7, 8, sparc_g0);
                        sparc_nop (code);
                        break;
                }
                case OP_CALL_HANDLER: 
                        mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
+                       /* This is a jump inside the method, so call_simple works even on V9 */
                        sparc_call_simple (code, 0);
                        sparc_nop (code);
                        break;
                case OP_LABEL:
                        ins->inst_c0 = (guint8*)code - cfg->native_code;
                        break;
-               case CEE_BR:
+               case OP_RELAXED_NOP:
+               case OP_NOP:
+               case OP_DUMMY_USE:
+               case OP_DUMMY_STORE:
+               case OP_NOT_REACHED:
+               case OP_NOT_NULL:
+                       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;
@@ -3071,18 +3233,37 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_CLT_UN:
                case OP_CGT:
                case OP_CGT_UN:
-                       //if (cfg->opt & MONO_OPT_CMOV) {
-                       if (0) {
+                       if (v64 && (cfg->opt & MONO_OPT_CMOV)) {
                                sparc_clr_reg (code, ins->dreg);
-                               sparc_movcc_imm (code, sparc_icc, opcode_to_sparc_cond (ins->opcode), 1, ins->dreg);
+                               sparc_movcc_imm (code, sparc_xcc, opcode_to_sparc_cond (ins->opcode), 1, ins->dreg);
                        }
                        else {
                                sparc_clr_reg (code, ins->dreg);
+#ifdef SPARCV9
+                               sparc_branchp (code, 1, opcode_to_sparc_cond (ins->opcode), DEFAULT_ICC, 0, 2);
+#else
                                sparc_branch (code, 1, opcode_to_sparc_cond (ins->opcode), 2);
+#endif
                                /* delay slot */
                                sparc_set (code, 1, ins->dreg);
                        }
                        break;
+               case OP_ICEQ:
+               case OP_ICLT:
+               case OP_ICLT_UN:
+               case OP_ICGT:
+               case OP_ICGT_UN:
+                   if (v64 && (cfg->opt & MONO_OPT_CMOV)) {
+                               sparc_clr_reg (code, ins->dreg);
+                               sparc_movcc_imm (code, sparc_icc, opcode_to_sparc_cond (ins->opcode), 1, ins->dreg);
+                   }
+                   else {
+                       sparc_clr_reg (code, ins->dreg);
+                       sparc_branchp (code, 1, opcode_to_sparc_cond (ins->opcode), sparc_icc_short, 0, 2);
+                       /* delay slot */
+                       sparc_set (code, 1, ins->dreg);
+                   }
+                   break;
                case OP_COND_EXC_EQ:
                case OP_COND_EXC_NE_UN:
                case OP_COND_EXC_LT:
@@ -3097,26 +3278,63 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_COND_EXC_NO:
                case OP_COND_EXC_C:
                case OP_COND_EXC_NC:
+               case OP_COND_EXC_IEQ:
+               case OP_COND_EXC_INE_UN:
+               case OP_COND_EXC_ILT:
+               case OP_COND_EXC_ILT_UN:
+               case OP_COND_EXC_IGT:
+               case OP_COND_EXC_IGT_UN:
+               case OP_COND_EXC_IGE:
+               case OP_COND_EXC_IGE_UN:
+               case OP_COND_EXC_ILE:
+               case OP_COND_EXC_ILE_UN:
+               case OP_COND_EXC_IOV:
+               case OP_COND_EXC_INO:
+               case OP_COND_EXC_IC:
+               case OP_COND_EXC_INC:
+#ifdef SPARCV9
+                       NOT_IMPLEMENTED;
+#else
                        EMIT_COND_SYSTEM_EXCEPTION (ins, opcode_to_sparc_cond (ins->opcode), ins->inst_p1);
+#endif
+                       break;
+               case OP_SPARC_COND_EXC_EQZ:
+                       EMIT_COND_SYSTEM_EXCEPTION_BPR (ins, brz, ins->inst_p1);
+                       break;
+               case OP_SPARC_COND_EXC_GEZ:
+                       EMIT_COND_SYSTEM_EXCEPTION_BPR (ins, brgez, ins->inst_p1);
+                       break;
+               case OP_SPARC_COND_EXC_GTZ:
+                       EMIT_COND_SYSTEM_EXCEPTION_BPR (ins, brgz, ins->inst_p1);
+                       break;
+               case OP_SPARC_COND_EXC_LEZ:
+                       EMIT_COND_SYSTEM_EXCEPTION_BPR (ins, brlez, ins->inst_p1);
+                       break;
+               case OP_SPARC_COND_EXC_LTZ:
+                       EMIT_COND_SYSTEM_EXCEPTION_BPR (ins, brlz, 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_SPARC_COND_EXC_NEZ:
+                       EMIT_COND_SYSTEM_EXCEPTION_BPR (ins, brnz, ins->inst_p1);
+                       break;
+
+               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: {
                        if (sparcv9)
                                EMIT_COND_BRANCH_PREDICTED (ins, opcode_to_sparc_cond (ins->opcode), 1, 1);
                        else
                                EMIT_COND_BRANCH (ins, opcode_to_sparc_cond (ins->opcode), 1, 1);
                        break;
                }
+
                case OP_SPARC_BRZ:
-                       /* We misuse the macro arguments */
                        EMIT_COND_BRANCH_BPR (ins, brz, 1, 1, 1);
                        break;
                case OP_SPARC_BRLEZ:
@@ -3138,21 +3356,30 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                /* floating point opcodes */
                case OP_R8CONST:
                        mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
+#ifdef SPARCV9
+                       sparc_set_template (code, sparc_o7);
+#else
                        sparc_sethi (code, 0, sparc_o7);
+#endif
                        sparc_lddf_imm (code, sparc_o7, 0, ins->dreg);
                        break;
                case OP_R4CONST:
                        mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
+#ifdef SPARCV9
+                       sparc_set_template (code, sparc_o7);
+#else
                        sparc_sethi (code, 0, sparc_o7);
-                       sparc_ldf_imm (code, sparc_o7, 0, ins->dreg);
+#endif
+                       sparc_ldf_imm (code, sparc_o7, 0, FP_SCRATCH_REG);
 
                        /* Extend to double */
-                       sparc_fstod (code, ins->dreg, ins->dreg);
+                       sparc_fstod (code, FP_SCRATCH_REG, ins->dreg);
                        break;
                case OP_STORER8_MEMBASE_REG:
                        if (!sparc_is_imm13 (ins->inst_offset + 4)) {
                                sparc_set (code, ins->inst_offset, sparc_o7);
-                               if (ins->inst_offset % 8) {
+                               /* SPARCV9 handles misaligned fp loads/stores */
+                               if (!v64 && (ins->inst_offset % 8)) {
                                        /* Misaligned */
                                        sparc_add (code, FALSE, ins->inst_destbasereg, sparc_o7, sparc_o7);
                                        sparc_stf (code, ins->sreg1, sparc_o7, sparc_g0);
@@ -3161,7 +3388,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                        sparc_stdf (code, ins->sreg1, ins->inst_destbasereg, sparc_o7);
                        }
                        else {
-                               if (ins->inst_offset % 8) {
+                               if (!v64 && (ins->inst_offset % 8)) {
                                        /* Misaligned */
                                        sparc_stf_imm (code, ins->sreg1, ins->inst_destbasereg, ins->inst_offset);
                                        sparc_stf_imm (code, ins->sreg1 + 1, ins->inst_destbasereg, ins->inst_offset + 4);
@@ -3170,62 +3397,113 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        }
                        break;
                case OP_LOADR8_MEMBASE:
-                       g_assert ((ins->inst_offset % 8) == 0);
                        EMIT_LOAD_MEMBASE (ins, lddf);
                        break;
                case OP_STORER4_MEMBASE_REG:
                        /* This requires a double->single conversion */
-                       sparc_fdtos (code, ins->sreg1, sparc_f0);
+                       sparc_fdtos (code, ins->sreg1, FP_SCRATCH_REG);
                        if (!sparc_is_imm13 (ins->inst_offset)) {
                                sparc_set (code, ins->inst_offset, sparc_o7);
-                               sparc_stf (code, sparc_f0, ins->inst_destbasereg, sparc_o7);
+                               sparc_stf (code, FP_SCRATCH_REG, ins->inst_destbasereg, sparc_o7);
                        }
                        else
-                               sparc_stf_imm (code, sparc_f0, ins->inst_destbasereg, ins->inst_offset);
-                       break;
-               case OP_LOADR4_MEMBASE:
-                       EMIT_LOAD_MEMBASE (ins, ldf);
-                       /* Extend to double */
-                       sparc_fstod (code, ins->dreg, ins->dreg);
+                               sparc_stf_imm (code, FP_SCRATCH_REG, ins->inst_destbasereg, ins->inst_offset);
                        break;
-               case OP_FMOVE:
-                       sparc_fmovs (code, ins->sreg1, ins->dreg);
-                       sparc_fmovs (code, ins->sreg1 + 1, ins->dreg + 1);
+               case OP_LOADR4_MEMBASE: {
+                       /* ldf needs a single precision register */
+                       int dreg = ins->dreg;
+                       ins->dreg = FP_SCRATCH_REG;
+                       EMIT_LOAD_MEMBASE (ins, ldf);
+                       ins->dreg = dreg;
+                       /* Extend to double */
+                       sparc_fstod (code, FP_SCRATCH_REG, ins->dreg);
                        break;
-               case CEE_CONV_R4: {
-                       guint32 offset = mono_spillvar_offset_float (cfg, 0);
-                       if (!sparc_is_imm13 (offset))
-                               NOT_IMPLEMENTED;
-                       sparc_st_imm (code, ins->sreg1, sparc_sp, offset);
-                       sparc_ldf_imm (code, sparc_sp, offset, sparc_f0);
-                       sparc_fitos (code, sparc_f0, sparc_f0);
-                       sparc_fstod (code, sparc_f0, ins->dreg);
+               }
+               case OP_ICONV_TO_R4: {
+                       gint32 offset = cfg->arch.float_spill_slot_offset;
+#ifdef SPARCV9
+                       if (!sparc_is_imm13 (offset)) {
+                               sparc_set (code, offset, sparc_o7);
+                               sparc_stx (code, ins->sreg1, sparc_sp, offset);
+                               sparc_lddf (code, sparc_sp, offset, FP_SCRATCH_REG);
+                       } else {
+                               sparc_stx_imm (code, ins->sreg1, sparc_sp, offset);
+                               sparc_lddf_imm (code, sparc_sp, offset, FP_SCRATCH_REG);
+                       }
+                       sparc_fxtos (code, FP_SCRATCH_REG, FP_SCRATCH_REG);
+#else
+                       if (!sparc_is_imm13 (offset)) {
+                               sparc_set (code, offset, sparc_o7);
+                               sparc_st (code, ins->sreg1, sparc_sp, sparc_o7);
+                               sparc_ldf (code, sparc_sp, sparc_o7, FP_SCRATCH_REG);
+                       } else {
+                               sparc_st_imm (code, ins->sreg1, sparc_sp, offset);
+                               sparc_ldf_imm (code, sparc_sp, offset, FP_SCRATCH_REG);
+                       }
+                       sparc_fitos (code, FP_SCRATCH_REG, FP_SCRATCH_REG);
+#endif
+                       sparc_fstod (code, FP_SCRATCH_REG, ins->dreg);
                        break;
                }
-               case CEE_CONV_R8: {
-                       guint32 offset = mono_spillvar_offset_float (cfg, 0);
-                       if (!sparc_is_imm13 (offset))
-                               NOT_IMPLEMENTED;
-                       sparc_st_imm (code, ins->sreg1, sparc_sp, offset);
-                       sparc_ldf_imm (code, sparc_sp, offset, sparc_f0);
-                       sparc_fitod (code, sparc_f0, ins->dreg);
+               case OP_ICONV_TO_R8: {
+                       gint32 offset = cfg->arch.float_spill_slot_offset;
+#ifdef SPARCV9
+                       if (!sparc_is_imm13 (offset)) {
+                               sparc_set (code, offset, sparc_o7);
+                               sparc_stx (code, ins->sreg1, sparc_sp, sparc_o7);
+                               sparc_lddf (code, sparc_sp, sparc_o7, FP_SCRATCH_REG);
+                       } else {
+                               sparc_stx_imm (code, ins->sreg1, sparc_sp, offset);
+                               sparc_lddf_imm (code, sparc_sp, offset, FP_SCRATCH_REG);
+                       }
+                       sparc_fxtod (code, FP_SCRATCH_REG, ins->dreg);
+#else
+                       if (!sparc_is_imm13 (offset)) {
+                               sparc_set (code, offset, sparc_o7);
+                               sparc_st (code, ins->sreg1, sparc_sp, sparc_o7);
+                               sparc_ldf (code, sparc_sp, sparc_o7, FP_SCRATCH_REG);
+                       } else {
+                               sparc_st_imm (code, ins->sreg1, sparc_sp, offset);
+                               sparc_ldf_imm (code, sparc_sp, offset, FP_SCRATCH_REG);
+                       }
+                       sparc_fitod (code, FP_SCRATCH_REG, ins->dreg);
+#endif
                        break;
                }
                case OP_FCONV_TO_I1:
                case OP_FCONV_TO_U1:
                case OP_FCONV_TO_I2:
                case OP_FCONV_TO_U2:
-               case OP_FCONV_TO_I4:
+#ifndef SPARCV9
                case OP_FCONV_TO_I:
-               case OP_FCONV_TO_U4:
-               case OP_FCONV_TO_U: {
-                       guint32 offset = mono_spillvar_offset_float (cfg, 0);
-                       if (!sparc_is_imm13 (offset))
-                               NOT_IMPLEMENTED;
-                       /* FIXME: Is having the same code for all of these ok ? */
-                       sparc_fdtoi (code, ins->sreg1, sparc_f0);
-                       sparc_stdf_imm (code, sparc_f0, sparc_sp, offset);
-                       sparc_ld_imm (code, sparc_sp, offset, ins->dreg);
+               case OP_FCONV_TO_U:
+#endif
+               case OP_FCONV_TO_I4:
+               case OP_FCONV_TO_U4: {
+                       gint32 offset = cfg->arch.float_spill_slot_offset;
+                       sparc_fdtoi (code, ins->sreg1, FP_SCRATCH_REG);
+                       if (!sparc_is_imm13 (offset)) {
+                               sparc_set (code, offset, sparc_o7);
+                               sparc_stdf (code, FP_SCRATCH_REG, sparc_sp, sparc_o7);
+                               sparc_ld (code, sparc_sp, sparc_o7, ins->dreg);
+                       } else {
+                               sparc_stdf_imm (code, FP_SCRATCH_REG, sparc_sp, offset);
+                               sparc_ld_imm (code, sparc_sp, offset, ins->dreg);
+                       }
+
+                       switch (ins->opcode) {
+                       case OP_FCONV_TO_I1:
+                       case OP_FCONV_TO_U1:
+                               sparc_and_imm (code, 0, ins->dreg, 0xff, ins->dreg);
+                               break;
+                       case OP_FCONV_TO_I2:
+                       case OP_FCONV_TO_U2:
+                               sparc_set (code, 0xffff, sparc_o7);
+                               sparc_and (code, 0, ins->dreg, sparc_o7, ins->dreg);
+                               break;
+                       default:
+                               break;
+                       }
                        break;
                }
                case OP_FCONV_TO_I8:
@@ -3233,16 +3511,22 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        /* Emulated */
                        g_assert_not_reached ();
                        break;
-               case CEE_CONV_R_UN:
-                       /* Emulated */
-                       g_assert_not_reached ();
+               case OP_FCONV_TO_R4:
+                       /* FIXME: Change precision ? */
+#ifdef SPARCV9
+                       sparc_fmovd (code, ins->sreg1, ins->dreg);
+#else
+                       sparc_fmovs (code, ins->sreg1, ins->dreg);
+                       sparc_fmovs (code, ins->sreg1 + 1, ins->dreg + 1);
+#endif
                        break;
                case OP_LCONV_TO_R_UN: { 
                        /* Emulated */
                        g_assert_not_reached ();
                        break;
                }
-               case OP_LCONV_TO_OVF_I: {
+               case OP_LCONV_TO_OVF_I:
+               case OP_LCONV_TO_OVF_I4_2: {
                        guint32 *br [3], *label [1];
 
                        /* 
@@ -3271,6 +3555,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        sparc_cmp_imm (code, ins->sreg2, -1);
                        br [2] = code;
                        sparc_branch (code, 1, sparc_bne, 0);
+                       sparc_nop (code);
                        sparc_patch (br [2], label [0]);
 
                        /* Ok */
@@ -3292,12 +3577,17 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        sparc_fdivd (code, ins->sreg1, ins->sreg2, ins->dreg);
                        break;          
                case OP_FNEG:
+#ifdef SPARCV9
+                       sparc_fnegd (code, ins->sreg1, ins->dreg);
+#else
+                       /* FIXME: why don't use fnegd ? */
                        sparc_fnegs (code, ins->sreg1, ins->dreg);
+#endif
                        break;          
                case OP_FREM:
-                       sparc_fdivd (code, ins->sreg1, ins->sreg2, sparc_f0);
-                       sparc_fmuld (code, ins->sreg2, sparc_f0, sparc_f0);
-                       sparc_fsubd (code, ins->sreg1, sparc_f0, ins->dreg);
+                       sparc_fdivd (code, ins->sreg1, ins->sreg2, FP_SCRATCH_REG);
+                       sparc_fmuld (code, ins->sreg2, FP_SCRATCH_REG, FP_SCRATCH_REG);
+                       sparc_fsubd (code, ins->sreg1, FP_SCRATCH_REG, ins->dreg);
                        break;
                case OP_FCOMPARE:
                        sparc_fcmpd (code, ins->sreg1, ins->sreg2);
@@ -3337,7 +3627,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        /* delay slot */
                        sparc_nop (code);
                        EMIT_FLOAT_COND_BRANCH (ins, sparc_fba, 1, 1);
-                       sparc_patch ((guint8*)p, (guint8*)code);
+                       sparc_patch (p, (guint8*)code);
                        break;
                }
                case OP_FBLE: {
@@ -3347,7 +3637,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        /* delay slot */
                        sparc_nop (code);
                        EMIT_FLOAT_COND_BRANCH (ins, sparc_fba, 1, 1);
-                       sparc_patch ((guint8*)p, (guint8*)code);
+                       sparc_patch (p, (guint8*)code);
                        break;
                }
                case OP_FBNE_UN:
@@ -3370,20 +3660,33 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        EMIT_FLOAT_COND_BRANCH (ins, sparc_fble, 1, 1);
                        EMIT_FLOAT_COND_BRANCH (ins, sparc_fbu, 1, 1);
                        break;
-               case CEE_CKFINITE: {
-                       guint32 offset = mono_spillvar_offset_float (cfg, 0);
-                       if (!sparc_is_imm13 (offset))
-                               NOT_IMPLEMENTED;
-                       sparc_stdf_imm (code, ins->sreg1, sparc_sp, offset);
-                       sparc_lduh_imm (code, sparc_sp, offset, sparc_o7);
+               case OP_CKFINITE: {
+                       gint32 offset = cfg->arch.float_spill_slot_offset;
+                       if (!sparc_is_imm13 (offset)) {
+                               sparc_set (code, offset, sparc_o7);
+                               sparc_stdf (code, ins->sreg1, sparc_sp, sparc_o7);
+                               sparc_lduh (code, sparc_sp, sparc_o7, sparc_o7);
+                       } else {
+                               sparc_stdf_imm (code, ins->sreg1, sparc_sp, offset);
+                               sparc_lduh_imm (code, sparc_sp, offset, sparc_o7);
+                       }
                        sparc_srl_imm (code, sparc_o7, 4, sparc_o7);
                        sparc_and_imm (code, FALSE, sparc_o7, 2047, sparc_o7);
                        sparc_cmp_imm (code, sparc_o7, 2047);
                        EMIT_COND_SYSTEM_EXCEPTION (ins, sparc_be, "ArithmeticException");
+#ifdef SPARCV9
+                       sparc_fmovd (code, ins->sreg1, ins->dreg);
+#else
                        sparc_fmovs (code, ins->sreg1, ins->dreg);
                        sparc_fmovs (code, ins->sreg1 + 1, ins->dreg + 1);
+#endif
                        break;
                }
+
+               case OP_MEMORY_BARRIER:
+                       sparc_membar (code, sparc_membar_all);
+                       break;
+
                default:
 #ifdef __GNUC__
                        g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
@@ -3402,8 +3705,6 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                cpos += max_len;
 
                last_ins = ins;
-               
-               ins = ins->next;
        }
 
        cfg->code_len = (guint8*)code - cfg->native_code;
@@ -3412,7 +3713,6 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 void
 mono_arch_register_lowlevel_calls (void)
 {
-       mono_register_jit_icall (mono_sparc_break, "mono_sparc_break", NULL, TRUE);
        mono_register_jit_icall (mono_arch_get_lmf_addr, "mono_arch_get_lmf_addr", NULL, TRUE);
 }
 
@@ -3424,77 +3724,94 @@ mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, Mono
        /* FIXME: Move part of this to arch independent code */
        for (patch_info = ji; patch_info; patch_info = patch_info->next) {
                unsigned char *ip = patch_info->ip.i + code;
-               const unsigned char *target = NULL;
+               gpointer target;
 
                target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
 
                switch (patch_info->type) {
+               case MONO_PATCH_INFO_NONE:
+                       continue;
                case MONO_PATCH_INFO_CLASS_INIT: {
-                       unsigned char *ip2 = ip;
+                       guint32 *ip2 = (guint32*)ip;
                        /* Might already been changed to a nop */
+#ifdef SPARCV9
+                       sparc_set_template (ip2, sparc_o7);
+                       sparc_jmpl (ip2, sparc_o7, sparc_g0, sparc_o7);
+#else
                        sparc_call_simple (ip2, 0);
+#endif
                        break;
                }
-               case MONO_PATCH_INFO_R4: {
-                       float *f = g_new0 (float, 1);
-                       *f = *(float*)patch_info->data.target;
-                       target = f;
-                       break;
-               }
-               case MONO_PATCH_INFO_R8: {
-                       double *d = g_new0 (double, 1);
-                       *d = *(double*)patch_info->data.target;
-                       target = d;                     
+               case MONO_PATCH_INFO_METHOD_JUMP: {
+                       guint32 *ip2 = (guint32*)ip;
+                       /* Might already been patched */
+                       sparc_set_template (ip2, sparc_o7);
                        break;
                }
                default:
                        break;
                }
-               sparc_patch (ip, target);
+               sparc_patch ((guint32*)ip, target);
        }
 }
 
-/*
- * Allow tracing to work with this interface (with an optional argument)
- */
-
-/*
- * This may be needed on some archs or for debugging support.
- */
-void
-mono_arch_instrument_mem_needs (MonoMethod *method, int *stack, int *code)
-{
-       /* no stack room needed now (may be needed for FASTCALL-trace support) */
-       *stack = 0;
-       /* split prolog-epilog requirements? */
-       *code = 256; /* max bytes needed: check this number */
-}
-
 void*
 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
 {
-       int stack, code_size;
+       int i;
        guint32 *code = (guint32*)p;
+       MonoMethodSignature *sig = mono_method_signature (cfg->method);
+       CallInfo *cinfo;
 
        /* Save registers to stack */
-       sparc_st_imm (code, sparc_i0, sparc_fp, 68);
-       sparc_st_imm (code, sparc_i1, sparc_fp, 72);
-       sparc_st_imm (code, sparc_i2, sparc_fp, 76);
-       sparc_st_imm (code, sparc_i3, sparc_fp, 80);
-       sparc_st_imm (code, sparc_i4, sparc_fp, 84);
-       sparc_st_imm (code, sparc_i5, sparc_fp, 88);
+       for (i = 0; i < 6; ++i)
+               sparc_sti_imm (code, sparc_i0 + i, sparc_fp, ARGS_OFFSET + (i * sizeof (gpointer)));
+
+       cinfo = get_call_info (cfg, sig, FALSE);
+
+       /* Save float regs on V9, since they are caller saved */
+       for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
+               ArgInfo *ainfo = cinfo->args + i;
+               gint32 stack_offset;
+
+               stack_offset = ainfo->offset + ARGS_OFFSET;
+
+               if (ainfo->storage == ArgInFloatReg) {
+                       if (!sparc_is_imm13 (stack_offset))
+                               NOT_IMPLEMENTED;
+                       sparc_stf_imm (code, ainfo->reg, sparc_fp, stack_offset);
+               }
+               else if (ainfo->storage == ArgInDoubleReg) {
+                       /* The offset is guaranteed to be aligned by the ABI rules */
+                       sparc_stdf_imm (code, ainfo->reg, sparc_fp, stack_offset);
+               }
+       }
 
        sparc_set (code, cfg->method, sparc_o0);
-       sparc_mov_reg_reg (code, sparc_fp, sparc_o1);
+       sparc_add_imm (code, FALSE, sparc_fp, MONO_SPARC_STACK_BIAS, sparc_o1);
 
        mono_add_patch_info (cfg, (guint8*)code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
-       sparc_sethi (code, 0, sparc_o7);
-       sparc_jmpl_imm (code, sparc_o7, 0, sparc_callsite);
-       sparc_nop (code);
+       EMIT_CALL ();
+
+       /* Restore float regs on V9 */
+       for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
+               ArgInfo *ainfo = cinfo->args + i;
+               gint32 stack_offset;
 
-       mono_arch_instrument_mem_needs (cfg->method, &stack, &code_size);
+               stack_offset = ainfo->offset + ARGS_OFFSET;
+
+               if (ainfo->storage == ArgInFloatReg) {
+                       if (!sparc_is_imm13 (stack_offset))
+                               NOT_IMPLEMENTED;
+                       sparc_ldf_imm (code, sparc_fp, stack_offset, ainfo->reg);
+               }
+               else if (ainfo->storage == ArgInDoubleReg) {
+                       /* The offset is guaranteed to be aligned by the ABI rules */
+                       sparc_lddf_imm (code, sparc_fp, stack_offset, ainfo->reg);
+               }
+       }
 
-       g_assert ((code - (guint32*)p) <= (code_size * 4));
+       g_free (cinfo);
 
        return code;
 }
@@ -3510,13 +3827,11 @@ enum {
 void*
 mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
 {
-       guchar *code = p;
+       guint32 *code = (guint32*)p;
        int save_mode = SAVE_NONE;
        MonoMethod *method = cfg->method;
-       int rtype = method->signature->ret->type;
-       
-handle_enum:
-       switch (rtype) {
+
+       switch (mono_type_get_underlying_type (mono_method_signature (method)->ret)->type) {
        case MONO_TYPE_VOID:
                /* special case string .ctor icall */
                if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
@@ -3526,17 +3841,17 @@ handle_enum:
                break;
        case MONO_TYPE_I8:
        case MONO_TYPE_U8:
+#ifdef SPARCV9
+               save_mode = SAVE_ONE;
+#else
                save_mode = SAVE_TWO;
+#endif
                break;
        case MONO_TYPE_R4:
        case MONO_TYPE_R8:
                save_mode = SAVE_FP;
                break;
        case MONO_TYPE_VALUETYPE:
-               if (method->signature->ret->data.klass->enumtype) {
-                       rtype = method->signature->ret->data.klass->enum_basetype->type;
-                       goto handle_enum;
-               }
                save_mode = SAVE_STRUCT;
                break;
        default:
@@ -3548,22 +3863,31 @@ handle_enum:
 
        switch (save_mode) {
        case SAVE_TWO:
+               /* V8 only */
                sparc_st_imm (code, sparc_i0, sparc_fp, 68);
                sparc_st_imm (code, sparc_i0, sparc_fp, 72);
                sparc_mov_reg_reg (code, sparc_i0, sparc_o1);
                sparc_mov_reg_reg (code, sparc_i1, sparc_o2);
                break;
        case SAVE_ONE:
-               sparc_st_imm (code, sparc_i0, sparc_fp, 68);
+               sparc_sti_imm (code, sparc_i0, sparc_fp, ARGS_OFFSET);
                sparc_mov_reg_reg (code, sparc_i0, sparc_o1);
                break;
        case SAVE_FP:
+#ifdef SPARCV9
+               sparc_stdf_imm (code, sparc_f0, sparc_fp, ARGS_OFFSET);
+#else
                sparc_stdf_imm (code, sparc_f0, sparc_fp, 72);
                sparc_ld_imm (code, sparc_fp, 72, sparc_o1);
-               sparc_ld_imm (code, sparc_fp, 72, sparc_o2);
+               sparc_ld_imm (code, sparc_fp, 72 + 4, sparc_o2);
+#endif
                break;
        case SAVE_STRUCT:
+#ifdef SPARCV9
+               sparc_mov_reg_reg (code, sparc_i0, sparc_o1);
+#else
                sparc_ld_imm (code, sparc_fp, 64, sparc_o1);
+#endif
                break;
        case SAVE_NONE:
        default:
@@ -3572,10 +3896,8 @@ handle_enum:
 
        sparc_set (code, cfg->method, sparc_o0);
 
-       mono_add_patch_info (cfg, (guint8*)code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
-       sparc_sethi (code, 0, sparc_o7);
-       sparc_jmpl_imm (code, sparc_o7, 0, sparc_callsite);
-       sparc_nop (code);
+       mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_ABS, func);
+       EMIT_CALL ();
 
        /* Restore result */
 
@@ -3585,10 +3907,10 @@ handle_enum:
                sparc_ld_imm (code, sparc_fp, 72, sparc_i0);
                break;
        case SAVE_ONE:
-               sparc_ld_imm (code, sparc_fp, 68, sparc_i0);
+               sparc_ldi_imm (code, sparc_fp, ARGS_OFFSET, sparc_i0);
                break;
        case SAVE_FP:
-               sparc_lddf_imm (code, sparc_fp, 72, sparc_f0);
+               sparc_lddf_imm (code, sparc_fp, ARGS_OFFSET, sparc_f0);
                break;
        case SAVE_NONE:
        default:
@@ -3598,63 +3920,36 @@ handle_enum:
        return code;
 }
 
-int
-mono_arch_max_epilog_size (MonoCompile *cfg)
-{
-       int exc_count = 0, max_epilog_size = 16 + 20*4;
-       MonoJumpInfo *patch_info;
-       
-       if (cfg->method->save_lmf)
-               max_epilog_size += 128;
-       
-       if (mono_jit_trace_calls != NULL)
-               max_epilog_size += 50;
-
-       if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
-               max_epilog_size += 50;
-
-       /* count the number of exception infos */
-     
-       for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
-               if (patch_info->type == MONO_PATCH_INFO_EXC)
-                       exc_count++;
-       }
-
-       /* 
-        * make sure we have enough space for exceptions
-        */
-       max_epilog_size += exc_count * 24;
-
-       return max_epilog_size;
-}
-
 guint8 *
 mono_arch_emit_prolog (MonoCompile *cfg)
 {
        MonoMethod *method = cfg->method;
        MonoMethodSignature *sig;
        MonoInst *inst;
-       guint8 *code;
+       guint32 *code;
        CallInfo *cinfo;
        guint32 i, offset;
 
        cfg->code_size = 256;
-       code = cfg->native_code = g_malloc (cfg->code_size);
+       cfg->native_code = g_malloc (cfg->code_size);
+       code = (guint32*)cfg->native_code;
 
        /* FIXME: Generate intermediate code instead */
 
        offset = cfg->stack_offset;
-       offset += 64; /* register save area */
+       offset += (16 * sizeof (gpointer)); /* register save area */
+#ifndef SPARCV9
        offset += 4; /* struct/union return pointer */
+#endif
 
        /* add parameter area size for called functions */
-       if (cfg->param_area < 24)
+       if (cfg->param_area < (6 * sizeof (gpointer)))
                /* Reserve space for the first 6 arguments even if it is unused */
-               offset += 24;
+               offset += 6 * sizeof (gpointer);
        else
                offset += cfg->param_area;
        
-       /* align the stack size to 8 bytes */
+       /* align the stack size */
        offset = ALIGN_TO (offset, MONO_ARCH_FRAME_ALIGNMENT);
 
        /*
@@ -3666,37 +3961,44 @@ mono_arch_emit_prolog (MonoCompile *cfg)
 
        cfg->stack_offset = offset;
 
+#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
+                       /* Perform stack touching */
+                       NOT_IMPLEMENTED;
+#endif
+
        if (!sparc_is_imm13 (- cfg->stack_offset)) {
                /* Can't use sparc_o7 here, since we're still in the caller's frame */
-               sparc_set (code, (- cfg->stack_offset), sparc_g1);
-               sparc_save (code, sparc_sp, sparc_g1, sparc_sp);
+               sparc_set (code, (- cfg->stack_offset), GP_SCRATCH_REG);
+               sparc_save (code, sparc_sp, GP_SCRATCH_REG, sparc_sp);
        }
        else
                sparc_save_imm (code, sparc_sp, - cfg->stack_offset, sparc_sp);
 
-       if (strstr (cfg->method->name, "test_marshal_struct")) {
-               mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_ABS, mono_sparc_break);
+/*
+       if (strstr (cfg->method->name, "foo")) {
+               mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_ABS, mono_sparc_break);
                sparc_call_simple (code, 0);
                sparc_nop (code);
        }
+*/
 
-       sig = method->signature;
+       sig = mono_method_signature (method);
 
-       cinfo = get_call_info (sig, FALSE);
+       cinfo = get_call_info (cfg, sig, FALSE);
 
        /* Keep in sync with emit_load_volatile_arguments */
        for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
                ArgInfo *ainfo = cinfo->args + i;
-               guint32 stack_offset;
+               gint32 stack_offset;
                MonoType *arg_type;
-               inst = cfg->varinfo [i];
+               inst = cfg->args [i];
 
                if (sig->hasthis && (i == 0))
-                       arg_type = mono_defaults.object_class;
+                       arg_type = &mono_defaults.object_class->byval_arg;
                else
                        arg_type = sig->params [i - sig->hasthis];
 
-               stack_offset = ainfo->offset + 68;
+               stack_offset = ainfo->offset + ARGS_OFFSET;
 
                /* Save the split arguments so they will reside entirely on the stack */
                if (ainfo->storage == ArgInSplitRegStack) {
@@ -3707,7 +4009,7 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                        sparc_st_imm (code, sparc_i5, inst->inst_basereg, stack_offset);
                }
 
-               if (!arg_type->byref && (arg_type->type == MONO_TYPE_R8)) {
+               if (!v64 && !arg_type->byref && (arg_type->type == MONO_TYPE_R8)) {
                        /* Save the argument to a dword aligned stack location */
                        /*
                         * stack_offset contains the offset of the argument on the stack.
@@ -3722,6 +4024,9 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                        }
                        else
                                if (ainfo->storage == ArgInSplitRegStack) {
+#ifdef SPARCV9
+                                       g_assert_not_reached ();
+#endif
                                        if (stack_offset != inst->inst_offset) {
                                                /* stack_offset is not dword aligned, so we need to make a copy */
                                                sparc_st_imm (code, sparc_i5, inst->inst_basereg, inst->inst_offset);
@@ -3731,6 +4036,9 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                                }
                        else
                                if (ainfo->storage == ArgOnStackPair) {
+#ifdef SPARCV9
+                                       g_assert_not_reached ();
+#endif
                                        if (stack_offset != inst->inst_offset) {
                                                /* stack_offset is not dword aligned, so we need to make a copy */
                                                sparc_ld_imm (code, sparc_fp, stack_offset, sparc_o7);
@@ -3747,22 +4055,46 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                                /* Argument in register, but need to be saved to stack */
                                if (!sparc_is_imm13 (stack_offset))
                                        NOT_IMPLEMENTED;
-                               if (stack_offset & 0x1)
+                               if ((stack_offset - ARGS_OFFSET) & 0x1)
                                        sparc_stb_imm (code, sparc_i0 + ainfo->reg, inst->inst_basereg, stack_offset);
                                else
-                                       if (stack_offset & 0x2)
+                                       if ((stack_offset - ARGS_OFFSET) & 0x2)
                                                sparc_sth_imm (code, sparc_i0 + ainfo->reg, inst->inst_basereg, stack_offset);
                                else
-                                       sparc_st_imm (code, sparc_i0 + ainfo->reg, inst->inst_basereg, stack_offset);
+                                       if ((stack_offset - ARGS_OFFSET) & 0x4)
+                                               sparc_st_imm (code, sparc_i0 + ainfo->reg, inst->inst_basereg, stack_offset);                           
+                                       else {
+                                               if (v64)
+                                                       sparc_stx_imm (code, sparc_i0 + ainfo->reg, inst->inst_basereg, stack_offset);
+                                               else
+                                                       sparc_st_imm (code, sparc_i0 + ainfo->reg, inst->inst_basereg, stack_offset);
+                                       }
                        }
                else
                        if ((ainfo->storage == ArgInIRegPair) && (inst->opcode != OP_REGVAR)) {
+#ifdef SPARCV9
+                               NOT_IMPLEMENTED;
+#endif
                                /* Argument in regpair, but need to be saved to stack */
                                if (!sparc_is_imm13 (inst->inst_offset + 4))
                                        NOT_IMPLEMENTED;
                                sparc_st_imm (code, sparc_i0 + ainfo->reg, inst->inst_basereg, inst->inst_offset);
                                sparc_st_imm (code, sparc_i0 + ainfo->reg + 1, inst->inst_basereg, inst->inst_offset + 4);                              
                        }
+               else if ((ainfo->storage == ArgInFloatReg) && (inst->opcode != OP_REGVAR)) {
+                               if (!sparc_is_imm13 (stack_offset))
+                                       NOT_IMPLEMENTED;
+                               sparc_stf_imm (code, ainfo->reg, inst->inst_basereg, inst->inst_offset);
+                               }
+                       else if ((ainfo->storage == ArgInDoubleReg) && (inst->opcode != OP_REGVAR)) {
+                               /* The offset is guaranteed to be aligned by the ABI rules */
+                               sparc_stdf_imm (code, ainfo->reg, inst->inst_basereg, inst->inst_offset);
+                       }
+                                       
+               if ((ainfo->storage == ArgInFloatReg) && (inst->opcode == OP_REGVAR)) {
+                       /* Need to move into the a double precision register */
+                       sparc_fstod (code, ainfo->reg, ainfo->reg - 1);
+               }
 
                if ((ainfo->storage == ArgInSplitRegStack) || (ainfo->storage == ArgOnStack))
                        if (inst->opcode == OP_REGVAR)
@@ -3773,89 +4105,247 @@ mono_arch_emit_prolog (MonoCompile *cfg)
        g_free (cinfo);
 
        if (cfg->method->save_lmf) {
-               gint32 lmf_offset = - cfg->arch.lmf_offset;
+               gint32 lmf_offset = STACK_BIAS - cfg->arch.lmf_offset;
 
                /* Save ip */
                mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_IP, NULL);
-               sparc_set (code, 0xfffffff, sparc_o7);
-               sparc_st_imm (code, sparc_o7, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ip));
+               sparc_set_template (code, sparc_o7);
+               sparc_sti_imm (code, sparc_o7, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ip));
                /* Save sp */
-               sparc_st_imm (code, sparc_sp, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, sp));
+               sparc_sti_imm (code, sparc_sp, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, sp));
                /* Save fp */
-               sparc_st_imm (code, sparc_fp, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp));
+               sparc_sti_imm (code, sparc_fp, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp));
                /* Save method */
                /* FIXME: add a relocation for this */
                sparc_set (code, cfg->method, sparc_o7);
-               sparc_st_imm (code, sparc_o7, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method));
+               sparc_sti_imm (code, sparc_o7, sparc_fp, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method));
 
                mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, 
                                                         (gpointer)"mono_arch_get_lmf_addr");           
-               sparc_call_simple (code, 0);
-               sparc_nop (code);
+               EMIT_CALL ();
 
-               code = (guint32*)mono_sparc_emit_save_lmf ((guint32*)code, lmf_offset);
+               code = (guint32*)mono_sparc_emit_save_lmf (code, lmf_offset);
        }
 
        if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
                code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
 
-       cfg->code_len = code - cfg->native_code;
+       cfg->code_len = (guint8*)code - cfg->native_code;
 
        g_assert (cfg->code_len <= cfg->code_size);
 
-       return code;
+       return (guint8*)code;
 }
 
 void
 mono_arch_emit_epilog (MonoCompile *cfg)
 {
-       MonoJumpInfo *patch_info;
        MonoMethod *method = cfg->method;
-       guint8 *code;
+       guint32 *code;
+       int can_fold = 0;
+       int max_epilog_size = 16 + 20 * 4;
+       
+       if (cfg->method->save_lmf)
+               max_epilog_size += 128;
+       
+       if (mono_jit_trace_calls != NULL)
+               max_epilog_size += 50;
+
+       if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
+               max_epilog_size += 50;
+
+       while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
+               cfg->code_size *= 2;
+               cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
+               mono_jit_stats.code_reallocs++;
+       }
 
-       code = cfg->native_code + cfg->code_len;
+       code = (guint32*)(cfg->native_code + cfg->code_len);
 
        if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
                code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
 
        if (cfg->method->save_lmf) {
-               gint32 lmf_offset = - cfg->arch.lmf_offset;
+               gint32 lmf_offset = STACK_BIAS - cfg->arch.lmf_offset;
 
                code = mono_sparc_emit_restore_lmf (code, lmf_offset);
        }
 
        /* 
-        * The sparc ABI requires that calls to functions which return a structure
+        * The V8 ABI requires that calls to functions which return a structure
         * return to %i7+12
         */
-       if (cfg->method->signature->pinvoke && MONO_TYPE_ISSTRUCT(cfg->method->signature->ret))
+       if (!v64 && mono_method_signature (cfg->method)->pinvoke && MONO_TYPE_ISSTRUCT(mono_method_signature (cfg->method)->ret))
                sparc_jmpl_imm (code, sparc_i7, 12, sparc_g0);
        else
                sparc_ret (code);
-       sparc_restore_imm (code, sparc_g0, 0, sparc_g0);
 
-       /* add code to raise exceptions */
+       /* Only fold last instruction into the restore if the exit block has an in count of 1
+          and the previous block hasn't been optimized away since it may have an in count > 1 */
+       if (cfg->bb_exit->in_count == 1 && cfg->bb_exit->in_bb[0]->native_offset != cfg->bb_exit->native_offset)
+               can_fold = 1;
+
+       /* 
+        * FIXME: The last instruction might have a branch pointing into it like in 
+        * int_ceq sparc_i0 <-
+        */
+       can_fold = 0;
+
+       /* Try folding last instruction into the restore */
+       if (can_fold && (sparc_inst_op (code [-2]) == 0x2) && (sparc_inst_op3 (code [-2]) == 0x2) && sparc_inst_imm (code [-2]) && (sparc_inst_rd (code [-2]) == sparc_i0)) {
+               /* or reg, imm, %i0 */
+               int reg = sparc_inst_rs1 (code [-2]);
+               int imm = (((gint32)(sparc_inst_imm13 (code [-2]))) << 19) >> 19;
+               code [-2] = code [-1];
+               code --;
+               sparc_restore_imm (code, reg, imm, sparc_o0);
+       }
+       else
+       if (can_fold && (sparc_inst_op (code [-2]) == 0x2) && (sparc_inst_op3 (code [-2]) == 0x2) && (!sparc_inst_imm (code [-2])) && (sparc_inst_rd (code [-2]) == sparc_i0)) {
+               /* or reg, reg, %i0 */
+               int reg1 = sparc_inst_rs1 (code [-2]);
+               int reg2 = sparc_inst_rs2 (code [-2]);
+               code [-2] = code [-1];
+               code --;
+               sparc_restore (code, reg1, reg2, sparc_o0);
+       }
+       else
+               sparc_restore_imm (code, sparc_g0, 0, sparc_g0);
+
+       cfg->code_len = (guint8*)code - cfg->native_code;
+
+       g_assert (cfg->code_len < cfg->code_size);
+
+}
+
+void
+mono_arch_emit_exceptions (MonoCompile *cfg)
+{
+       MonoJumpInfo *patch_info;
+       guint32 *code;
+       int nthrows = 0, i;
+       int exc_count = 0;
+       guint32 code_size;
+       MonoClass *exc_classes [16];
+       guint8 *exc_throw_start [16], *exc_throw_end [16];
+
+       /* Compute needed space */
+       for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
+               if (patch_info->type == MONO_PATCH_INFO_EXC)
+                       exc_count++;
+       }
+     
+       /* 
+        * make sure we have enough space for exceptions
+        */
+#ifdef SPARCV9
+       code_size = exc_count * (20 * 4);
+#else
+       code_size = exc_count * 24;
+#endif
+
+       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);
+               mono_jit_stats.code_reallocs++;
+       }
+
+       code = (guint32*)(cfg->native_code + cfg->code_len);
+
        for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
                switch (patch_info->type) {
-               case MONO_PATCH_INFO_EXC:
-                       sparc_patch (cfg->native_code + patch_info->ip.i, code);
-                       mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_EXC_NAME, patch_info->data.target);
-                       sparc_set (code, 0xffffff, sparc_o0);
-                       mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_METHOD_REL, (gpointer)patch_info->ip.i);
-                       sparc_set (code, 0xffffff, sparc_o1);
-                       patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
-                       patch_info->data.name = "mono_arch_throw_exception_by_name";
-                       patch_info->ip.i = code - cfg->native_code;
-                       sparc_call_simple (code, 0);
-                       sparc_nop (code);
+               case MONO_PATCH_INFO_EXC: {
+                       MonoClass *exc_class;
+                       guint32 *buf, *buf2;
+                       guint32 throw_ip, type_idx;
+                       gint32 disp;
+
+                       sparc_patch ((guint32*)(cfg->native_code + patch_info->ip.i), code);
+
+                       exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
+                       g_assert (exc_class);
+                       type_idx = exc_class->type_token - MONO_TOKEN_TYPE_DEF;
+                       throw_ip = patch_info->ip.i;
+
+                       /* Find a throw sequence for the same exception class */
+                       for (i = 0; i < nthrows; ++i)
+                               if (exc_classes [i] == exc_class)
+                                       break;
+
+                       if (i < nthrows) {
+                               guint32 throw_offset = (((guint8*)exc_throw_end [i] - cfg->native_code) - throw_ip) >> 2;
+                               if (!sparc_is_imm13 (throw_offset))
+                                       sparc_set32 (code, throw_offset, sparc_o1);
+
+                               disp = (exc_throw_start [i] - (guint8*)code) >> 2;
+                               g_assert (sparc_is_imm22 (disp));
+                               sparc_branch (code, 0, sparc_ba, disp);
+                               if (sparc_is_imm13 (throw_offset))
+                                       sparc_set32 (code, throw_offset, sparc_o1);
+                               else
+                                       sparc_nop (code);
+                               patch_info->type = MONO_PATCH_INFO_NONE;
+                       }
+                       else {
+                               /* Emit the template for setting o1 */
+                               buf = code;
+                               if (sparc_is_imm13 (((((guint8*)code - cfg->native_code) - throw_ip) >> 2) - 8))
+                                       /* Can use a short form */
+                                       sparc_nop (code);
+                               else
+                                       sparc_set_template (code, sparc_o1);
+                               buf2 = code;
+
+                               if (nthrows < 16) {
+                                       exc_classes [nthrows] = exc_class;
+                                       exc_throw_start [nthrows] = (guint8*)code;
+                               }
+
+                               /*
+                               mono_add_patch_info (cfg, (guint8*)code - cfg->native_code, MONO_PATCH_INFO_ABS, mono_sparc_break);
+                               EMIT_CALL();
+                               */
+
+                               /* first arg = type token */
+                               /* Pass the type index to reduce the size of the sparc_set */
+                               if (!sparc_is_imm13 (type_idx))
+                                       sparc_set32 (code, type_idx, sparc_o0);
+
+                               /* second arg = offset between the throw ip and the current ip */
+                               /* On sparc, the saved ip points to the call instruction */
+                               disp = (((guint8*)code - cfg->native_code) - throw_ip) >> 2;
+                               sparc_set32 (buf, disp, sparc_o1);
+                               while (buf < buf2)
+                                       sparc_nop (buf);
+
+                               if (nthrows < 16) {
+                                       exc_throw_end [nthrows] = (guint8*)code;
+                                       nthrows ++;
+                               }
+
+                               patch_info->data.name = "mono_arch_throw_corlib_exception";
+                               patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
+                               patch_info->ip.i = (guint8*)code - cfg->native_code;
+
+                               EMIT_CALL ();
+
+                               if (sparc_is_imm13 (type_idx)) {
+                                       /* Put it into the delay slot */
+                                       code --;
+                                       buf = code;
+                                       sparc_set32 (code, type_idx, sparc_o0);
+                                       g_assert (code - buf == 1);
+                               }
+                       }
                        break;
+               }
                default:
                        /* do nothing */
                        break;
                }
        }
 
-       cfg->code_len = code - cfg->native_code;
+       cfg->code_len = (guint8*)code - cfg->native_code;
 
        g_assert (cfg->code_len < cfg->code_size);
 
@@ -3883,34 +4373,20 @@ mono_arch_get_lmf_addr (void)
 #endif
 }
 
-void
-mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
-{
 #ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
-#ifdef __linux__
-       struct sigaltstack sa;
-#else
-       stack_t         sigstk;
-#endif
-       printf ("SIGALT!\n");
-       /* Setup an alternate signal stack */
-       tls->signal_stack = g_malloc (SIGNAL_STACK_SIZE);
-       tls->signal_stack_size = SIGNAL_STACK_SIZE;
-
-#ifdef __linux__
-       sa.ss_sp = tls->signal_stack;
-       sa.ss_size = SIGNAL_STACK_SIZE;
-       sa.ss_flags = 0;
-       g_assert (sigaltstack (&sa, NULL) == 0);
-#else
-       sigstk.ss_sp = tls->signal_stack;
-       sigstk.ss_size = SIGNAL_STACK_SIZE;
-       sigstk.ss_flags = 0;
-       g_assert (sigaltstack (&sigstk, NULL) == 0);
-#endif
+
+/*
+ * There seems to be no way to determine stack boundaries under solaris,
+ * so it's not possible to determine whenever a SIGSEGV is caused by stack
+ * overflow or not.
+ */
+#error "--with-sigaltstack=yes not supported on solaris"
+
 #endif
 
+void
+mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
+{
        if (!lmf_addr_key_inited) {
                int res;
 
@@ -3937,30 +4413,12 @@ mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
 {
 }
 
-void
-mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
+MonoInst*
+mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
 {
-       /* add the this argument */
-       if (this_reg != -1) {
-               MonoInst *this;
-               MONO_INST_NEW (cfg, this, OP_SETREG);
-               this->type = this_type;
-               this->sreg1 = this_reg;
-               this->dreg = sparc_o0;
-               mono_bblock_add_inst (cfg->cbb, this);
-       }
-
-       if (vt_reg != -1) {
-               /* Set the 'struct/union return pointer' location on the stack */
-               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI4_MEMBASE_REG, sparc_sp, 64, vt_reg);
-       }
-}
+       MonoInst *ins = NULL;
 
-
-gint
-mono_arch_get_opcode_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
-{
-       return -1;
+       return ins;
 }
 
 /*
@@ -3981,23 +4439,22 @@ mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJit
        CallInfo *cinfo;
        ArgInfo *ainfo;
 
-       cinfo = get_call_info (csig, FALSE);
+       cinfo = get_call_info (NULL, csig, FALSE);
 
        if (csig->hasthis) {
                ainfo = &cinfo->args [0];
-               arg_info [0].offset = 68 + ainfo->offset;
+               arg_info [0].offset = ARGS_OFFSET - MONO_SPARC_STACK_BIAS + ainfo->offset;
        }
 
        for (k = 0; k < param_count; k++) {
                ainfo = &cinfo->args [k + csig->hasthis];
 
-               arg_info [k + 1].offset = 68 + ainfo->offset;
+               arg_info [k + 1].offset = ARGS_OFFSET - MONO_SPARC_STACK_BIAS + ainfo->offset;
                arg_info [k + 1].size = mono_type_size (csig->params [k], &align);
        }
 
        g_free (cinfo);
 
-       /* FIXME: */
        return 0;
 }
 
@@ -4006,3 +4463,20 @@ mono_arch_print_tree (MonoInst *tree, int arity)
 {
        return 0;
 }
+
+MonoInst* mono_arch_get_domain_intrinsic (MonoCompile* cfg)
+{
+       return NULL;
+}
+
+MonoInst* mono_arch_get_thread_intrinsic (MonoCompile* cfg)
+{
+       return NULL;
+}
+
+gpointer
+mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
+{
+       /* FIXME: implement */
+       g_assert_not_reached ();
+}