Merge pull request #601 from knocte/sock_improvements
[mono.git] / mono / mini / mini-codegen.c
index c122f9d68bf8f26456624d82c24c3015259c668f..8ef772a1b14c82c2c139ba05f0ea329e46bc9b16 100644 (file)
@@ -14,6 +14,7 @@
 #include <mono/metadata/debug-helpers.h>
 #include <mono/metadata/threads.h>
 #include <mono/metadata/profiler-private.h>
+#include <mono/metadata/mempool-internals.h>
 #include <mono/utils/mono-math.h>
 
 #include "mini.h"
 #define MONO_ARCH_CALLEE_XREGS 0
 
 #endif
+
+#define MONO_ARCH_BANK_MIRRORED -2
+
+#ifdef MONO_ARCH_USE_SHARED_FP_SIMD_BANK
+
+#ifndef MONO_ARCH_NEED_SIMD_BANK
+#error "MONO_ARCH_USE_SHARED_FP_SIMD_BANK needs MONO_ARCH_NEED_SIMD_BANK to work"
+#endif
+
+#define get_mirrored_bank(bank) (((bank) == MONO_REG_SIMD ) ? MONO_REG_DOUBLE : (((bank) == MONO_REG_DOUBLE ) ? MONO_REG_SIMD : -1))
+
+#define is_hreg_mirrored(rs, bank, hreg) ((rs)->symbolic [(bank)] [(hreg)] == MONO_ARCH_BANK_MIRRORED)
+
+
+#else
+
+
+#define get_mirrored_bank(bank) (-1)
+
+#define is_hreg_mirrored(rs, bank, hreg) (0)
+
+#endif
+
+
+/* If the bank is mirrored return the true logical bank that the register in the
+ * physical register bank is allocated to.
+ */
+static inline int translate_bank (MonoRegState *rs, int bank, int hreg) {
+       return is_hreg_mirrored (rs, bank, hreg) ? get_mirrored_bank (bank) : bank;
+}
+
 /*
  * Every hardware register belongs to a register type or register bank. bank 0 
  * contains the int registers, bank 1 contains the fp registers.
 static const int regbank_size [] = {
        MONO_MAX_IREGS,
        MONO_MAX_FREGS,
+       MONO_MAX_IREGS,
+       MONO_MAX_IREGS,
        MONO_MAX_XREGS
 };
 
 static const int regbank_load_ops [] = { 
-       OP_LOAD_MEMBASE,
+       OP_LOADR_MEMBASE,
        OP_LOADR8_MEMBASE,
+       OP_LOADR_MEMBASE,
+       OP_LOADR_MEMBASE,
        OP_LOADX_MEMBASE
 };
 
 static const int regbank_store_ops [] = { 
-       OP_STORE_MEMBASE_REG,
+       OP_STORER_MEMBASE_REG,
        OP_STORER8_MEMBASE_REG,
+       OP_STORER_MEMBASE_REG,
+       OP_STORER_MEMBASE_REG,
        OP_STOREX_MEMBASE
 };
 
 static const int regbank_move_ops [] = { 
        OP_MOVE,
        OP_FMOVE,
+       OP_MOVE,
+       OP_MOVE,
        OP_XMOVE
 };
 
@@ -63,62 +104,56 @@ static const int regbank_move_ops [] = {
 static const regmask_t regbank_callee_saved_regs [] = {
        MONO_ARCH_CALLEE_SAVED_REGS,
        MONO_ARCH_CALLEE_SAVED_FREGS,
+       MONO_ARCH_CALLEE_SAVED_REGS,
+       MONO_ARCH_CALLEE_SAVED_REGS,
        MONO_ARCH_CALLEE_SAVED_XREGS,
 };
 
 static const regmask_t regbank_callee_regs [] = {
        MONO_ARCH_CALLEE_REGS,
        MONO_ARCH_CALLEE_FREGS,
+       MONO_ARCH_CALLEE_REGS,
+       MONO_ARCH_CALLEE_REGS,
        MONO_ARCH_CALLEE_XREGS,
 };
 
 static const int regbank_spill_var_size[] = {
-       sizeof (gpointer),
+       sizeof (mgreg_t),
        sizeof (double),
+       sizeof (mgreg_t),
+       sizeof (mgreg_t),
        16 /*FIXME make this a constant. Maybe MONO_ARCH_SIMD_VECTOR_SIZE? */
 };
 
 #define DEBUG(a) MINI_DEBUG(cfg->verbose_level, 3, a;)
 
-static inline GSList*
-g_slist_append_mempool (MonoMemPool *mp, GSList *list, gpointer data)
-{
-       GSList *new_list;
-       GSList *last;
-       
-       new_list = mono_mempool_alloc (mp, sizeof (GSList));
-       new_list->data = data;
-       new_list->next = NULL;
-       
-       if (list) {
-               last = list;
-               while (last->next)
-                       last = last->next;
-               last->next = new_list;
-               
-               return list;
-       } else
-               return new_list;
-}
-
 static inline void
 mono_regstate_assign (MonoRegState *rs)
 {
+#ifdef MONO_ARCH_USE_SHARED_FP_SIMD_BANK
+       /* The regalloc may fail if fp and simd logical regbanks share the same physical reg bank and
+        * if the values here are not the same.
+        */
+       g_assert(regbank_callee_regs [MONO_REG_SIMD] == regbank_callee_regs [MONO_REG_DOUBLE]);
+       g_assert(regbank_callee_saved_regs [MONO_REG_SIMD] == regbank_callee_saved_regs [MONO_REG_DOUBLE]);
+       g_assert(regbank_size [MONO_REG_SIMD] == regbank_size [MONO_REG_DOUBLE]);
+#endif
+
        if (rs->next_vreg > rs->vassign_size) {
                g_free (rs->vassign);
                rs->vassign_size = MAX (rs->next_vreg, 256);
-               rs->vassign = g_malloc (rs->vassign_size * sizeof (int));
+               rs->vassign = g_malloc (rs->vassign_size * sizeof (gint32));
        }
 
        memset (rs->isymbolic, 0, MONO_MAX_IREGS * sizeof (rs->isymbolic [0]));
        memset (rs->fsymbolic, 0, MONO_MAX_FREGS * sizeof (rs->fsymbolic [0]));
 
-       rs->symbolic [0] = rs->isymbolic;
-       rs->symbolic [1] = rs->fsymbolic;
+       rs->symbolic [MONO_REG_INT] = rs->isymbolic;
+       rs->symbolic [MONO_REG_DOUBLE] = rs->fsymbolic;
 
 #ifdef MONO_ARCH_NEED_SIMD_BANK
        memset (rs->xsymbolic, 0, MONO_MAX_XREGS * sizeof (rs->xsymbolic [0]));
-       rs->symbolic [2] = rs->xsymbolic;
+       rs->symbolic [MONO_REG_SIMD] = rs->xsymbolic;
 #endif
 }
 
@@ -166,10 +201,17 @@ static inline int
 mono_regstate_alloc_general (MonoRegState *rs, regmask_t allow, int bank)
 {
        int i;
+       int mirrored_bank;
        regmask_t mask = allow & rs->free_mask [bank];
        for (i = 0; i < regbank_size [bank]; ++i) {
                if (mask & ((regmask_t)1 << i)) {
                        rs->free_mask [bank] &= ~ ((regmask_t)1 << i);
+
+                       mirrored_bank = get_mirrored_bank (bank);
+                       if (mirrored_bank == -1)
+                               return i;
+
+                       rs->free_mask [mirrored_bank] = rs->free_mask [bank];
                        return i;
                }
        }
@@ -179,9 +221,17 @@ mono_regstate_alloc_general (MonoRegState *rs, regmask_t allow, int bank)
 static inline void
 mono_regstate_free_general (MonoRegState *rs, int reg, int bank)
 {
+       int mirrored_bank;
+
        if (reg >= 0) {
                rs->free_mask [bank] |= (regmask_t)1 << reg;
                rs->symbolic [bank][reg] = 0;
+
+               mirrored_bank = get_mirrored_bank (bank);
+               if (mirrored_bank == -1)
+                       return;
+               rs->free_mask [mirrored_bank] = rs->free_mask [bank];
+               rs->symbolic [mirrored_bank][reg] = 0;
        }
 }
 
@@ -190,10 +240,12 @@ mono_regname_full (int reg, int bank)
 {
        if (G_UNLIKELY (bank)) {
 #if MONO_ARCH_NEED_SIMD_BANK
-               if (bank == 2)
+               if (bank == MONO_REG_SIMD)
                        return mono_arch_xregname (reg);
 #endif
-               g_assert (bank == 1);
+               if (bank == MONO_REG_INT_REF || bank == MONO_REG_INT_MP)
+                       return mono_arch_regname (reg);
+               g_assert (bank == MONO_REG_DOUBLE);
                return mono_arch_fregname (reg);
        } else {
                return mono_arch_regname (reg);
@@ -219,6 +271,17 @@ mono_call_inst_add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, int vreg, i
        }
 }
 
+/*
+ * mono_call_inst_add_outarg_vt:
+ *
+ *   Register OUTARG_VT as belonging to CALL.
+ */
+void
+mono_call_inst_add_outarg_vt (MonoCompile *cfg, MonoCallInst *call, MonoInst *outarg_vt)
+{
+       call->outarg_vts = g_slist_append_mempool (cfg->mempool, call->outarg_vts, outarg_vt);
+}
+
 static void
 resize_spill_info (MonoCompile *cfg, int bank)
 {
@@ -230,7 +293,7 @@ resize_spill_info (MonoCompile *cfg, int bank)
 
        g_assert (bank < MONO_NUM_REGBANKS);
 
-       new_info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo) * new_len);
+       new_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoSpillInfo) * new_len);
        if (orig_info)
                memcpy (new_info, orig_info, sizeof (MonoSpillInfo) * orig_len);
        for (i = orig_len; i < new_len; ++i)
@@ -260,14 +323,14 @@ mono_spillvar_offset (MonoCompile *cfg, int spillvar, int bank)
         */
        info = &cfg->spill_info [bank][spillvar];
        if (info->offset == -1) {
-               cfg->stack_offset += sizeof (gpointer) - 1;
-               cfg->stack_offset &= ~(sizeof (gpointer) - 1);
+               cfg->stack_offset += sizeof (mgreg_t) - 1;
+               cfg->stack_offset &= ~(sizeof (mgreg_t) - 1);
 
                g_assert (bank < MONO_NUM_REGBANKS);
                if (G_UNLIKELY (bank))
                        size = regbank_spill_var_size [bank];
                else
-                       size = sizeof (gpointer);
+                       size = sizeof (mgreg_t);
 
                if (cfg->flags & MONO_CFG_HAS_SPILLUP) {
                        cfg->stack_offset += size - 1;
@@ -304,8 +367,9 @@ mono_spillvar_offset (MonoCompile *cfg, int spillvar, int bank)
 
 #define reg_is_fp(desc) (MONO_ARCH_INST_IS_FLOAT (desc))
 #define dreg_is_fp(spec)  (MONO_ARCH_INST_IS_FLOAT (spec [MONO_INST_DEST]))
-#define sreg1_is_fp(spec) (MONO_ARCH_INST_IS_FLOAT (spec [MONO_INST_SRC1]))
-#define sreg2_is_fp(spec) (MONO_ARCH_INST_IS_FLOAT (spec [MONO_INST_SRC2]))
+#define sreg_is_fp(n,spec) (MONO_ARCH_INST_IS_FLOAT (spec [MONO_INST_SRC1+(n)]))
+#define sreg1_is_fp(spec) sreg_is_fp (0,(spec))
+#define sreg2_is_fp(spec) sreg_is_fp (1,(spec))
 
 #define reg_is_simd(desc) ((desc) == 'x') 
 
@@ -319,12 +383,14 @@ mono_spillvar_offset (MonoCompile *cfg, int spillvar, int bank)
 
 #endif
 
-#define sreg1_bank(spec) reg_bank ((spec)[MONO_INST_SRC1])
-#define sreg2_bank(spec) reg_bank ((spec)[MONO_INST_SRC2])
+#define sreg_bank(n,spec) reg_bank ((spec)[MONO_INST_SRC1+(n)])
+#define sreg1_bank(spec) sreg_bank (0, (spec))
+#define sreg2_bank(spec) sreg_bank (1, (spec))
 #define dreg_bank(spec) reg_bank ((spec)[MONO_INST_DEST])
 
-#define sreg1_bank_ins(ins) sreg1_bank (ins_get_spec ((ins)->opcode))
-#define sreg2_bank_ins(ins) sreg2_bank (ins_get_spec ((ins)->opcode))
+#define sreg_bank_ins(n,ins) sreg_bank ((n), ins_get_spec ((ins)->opcode))
+#define sreg1_bank_ins(ins) sreg_bank_ins (0, (ins))
+#define sreg2_bank_ins(ins) sreg_bank_ins (1, (ins))
 #define dreg_bank_ins(ins) dreg_bank (ins_get_spec ((ins)->opcode))
 
 #define regpair_reg2_mask(desc,hreg1) ((MONO_ARCH_INST_REGPAIR_REG2 (desc,hreg1) != -1) ? (regmask (MONO_ARCH_INST_REGPAIR_REG2 (desc,hreg1))) : MONO_ARCH_CALLEE_REGS)
@@ -344,10 +410,43 @@ typedef struct {
 } RegTrack;
 
 #ifndef DISABLE_LOGGING
+
+static const char* const patch_info_str[] = {
+#define PATCH_INFO(a,b) "" #a,
+#include "patch-info.h"
+#undef PATCH_INFO
+};
+
+void
+mono_print_ji (const MonoJumpInfo *ji)
+{
+       switch (ji->type) {
+       case MONO_PATCH_INFO_RGCTX_FETCH: {
+               MonoJumpInfoRgctxEntry *entry = ji->data.rgctx_entry;
+
+               printf ("[RGCTX_FETCH ");
+               mono_print_ji (entry->data);
+               printf (" - %s]", mono_rgctx_info_type_to_str (entry->info_type));
+               break;
+       }
+       case MONO_PATCH_INFO_METHODCONST: {
+               char *s = mono_method_full_name (ji->data.method, TRUE);
+               printf ("[METHODCONST - %s]", s);
+               g_free (s);
+               break;
+       }
+       default:
+               printf ("[%s]", patch_info_str [ji->type]);
+               break;
+       }
+}
+
 void
 mono_print_ins_index (int i, MonoInst *ins)
 {
        const char *spec = ins_get_spec (ins->opcode);
+       int num_sregs, j;
+       int sregs [MONO_MAX_SRC_REGS];
 
        if (i != -1)
                printf ("\t%-2d %s", i, mono_inst_name (ins->opcode));
@@ -361,6 +460,8 @@ mono_print_ins_index (int i, MonoInst *ins)
                        printf (" R%d", ins->sreg1);
                if (ins->sreg2 != -1)
                        printf (" R%d", ins->sreg2);
+               if (ins->sreg3 != -1)
+                       printf (" R%d", ins->sreg3);
 
                switch (ins->opcode) {
                case OP_LBNE_UN:
@@ -373,14 +474,14 @@ mono_print_ins_index (int i, MonoInst *ins)
                case OP_LBGE_UN:
                case OP_LBLE:
                case OP_LBLE_UN:
-                       if (!(ins->flags & MONO_INST_BRLABEL)) {
-                               if (!ins->inst_false_bb)
-                                       printf (" [B%d]", ins->inst_true_bb->block_num);
-                               else
-                                       printf (" [B%dB%d]", ins->inst_true_bb->block_num, ins->inst_false_bb->block_num);
-                       }
+                       if (!ins->inst_false_bb)
+                               printf (" [B%d]", ins->inst_true_bb->block_num);
+                       else
+                               printf (" [B%dB%d]", ins->inst_true_bb->block_num, ins->inst_false_bb->block_num);
                        break;
                case OP_PHI:
+               case OP_VPHI:
+               case OP_XPHI:
                case OP_FPHI: {
                        int i;
                        printf (" [%d (", (int)ins->inst_c0);
@@ -398,6 +499,7 @@ mono_print_ins_index (int i, MonoInst *ins)
                        break;
                case OP_REGOFFSET:
                        printf (" + 0x%lx", (long)ins->inst_offset);
+                       break;
                default:
                        break;
                }
@@ -438,19 +540,20 @@ mono_print_ins_index (int i, MonoInst *ins)
                else
                        printf (" %s", mono_regname_full (ins->sreg1, bank));
        }
-       if (spec [MONO_INST_SRC2]) {
-               int bank = sreg2_bank (spec);
-               if (is_soft_reg (ins->sreg2, bank))
-                       printf (" R%d", ins->sreg2);
+       num_sregs = mono_inst_get_src_registers (ins, sregs);
+       for (j = 1; j < num_sregs; ++j) {
+               int bank = sreg_bank (j, spec);
+               if (is_soft_reg (sregs [j], bank))
+                       printf (" R%d", sregs [j]);
                else
-                       printf (" %s", mono_regname_full (ins->sreg2, bank));
+                       printf (" %s", mono_regname_full (sregs [j], bank));
        }
 
        switch (ins->opcode) {
        case OP_ICONST:
                printf (" [%d]", (int)ins->inst_c0);
                break;
-#if defined(__i386__) || defined(__x86_64__)
+#if defined(TARGET_X86) || defined(TARGET_AMD64)
        case OP_X86_PUSH_IMM:
 #endif
        case OP_ICOMPARE_IMM:
@@ -460,6 +563,7 @@ mono_print_ins_index (int i, MonoInst *ins)
        case OP_IAND_IMM:
        case OP_IOR_IMM:
        case OP_IXOR_IMM:
+       case OP_SUB_IMM:
                printf (" [%d]", (int)ins->inst_imm);
                break;
        case OP_ADD_IMM:
@@ -475,8 +579,6 @@ mono_print_ins_index (int i, MonoInst *ins)
        case OP_R4CONST:
                printf (" [%f]", *(float*)ins->inst_p0);
                break;
-       case CEE_CALL:
-       case CEE_CALLVIRT:
        case OP_CALL:
        case OP_CALL_MEMBASE:
        case OP_CALL_REG:
@@ -492,7 +594,9 @@ mono_print_ins_index (int i, MonoInst *ins)
        case OP_VCALL2_REG:
        case OP_VCALL2_MEMBASE:
        case OP_VOIDCALL:
-       case OP_VOIDCALLVIRT: {
+       case OP_VOIDCALL_MEMBASE:
+       case OP_VOIDCALLVIRT:
+       case OP_TAILCALL: {
                MonoCallInst *call = (MonoCallInst*)ins;
                GSList *list;
 
@@ -509,6 +613,11 @@ mono_print_ins_index (int i, MonoInst *ins)
                        char *full_name = mono_method_full_name (call->method, TRUE);
                        printf (" [%s]", full_name);
                        g_free (full_name);
+               } else if (call->fptr_is_patch) {
+                       MonoJumpInfo *ji = (MonoJumpInfo*)call->fptr;
+
+                       printf (" ");
+                       mono_print_ji (ji);
                } else if (call->fptr) {
                        MonoJitICallInfo *info = mono_find_jit_icall_by_addr (call->fptr);
                        if (info)
@@ -534,16 +643,6 @@ mono_print_ins_index (int i, MonoInst *ins)
        case OP_CALL_HANDLER:
                printf (" [B%d]", ins->inst_target_bb->block_num);
                break;
-       case CEE_BNE_UN:
-       case CEE_BEQ:
-       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_IBNE_UN:
        case OP_IBEQ:
        case OP_IBLT:
@@ -564,12 +663,19 @@ mono_print_ins_index (int i, MonoInst *ins)
        case OP_LBGE_UN:
        case OP_LBLE:
        case OP_LBLE_UN:
-               if (!(ins->flags & MONO_INST_BRLABEL)) {
-                       if (!ins->inst_false_bb)
-                               printf (" [B%d]", ins->inst_true_bb->block_num);
-                       else
-                               printf (" [B%dB%d]", ins->inst_true_bb->block_num, ins->inst_false_bb->block_num);
-               }
+               if (!ins->inst_false_bb)
+                       printf (" [B%d]", ins->inst_true_bb->block_num);
+               else
+                       printf (" [B%dB%d]", ins->inst_true_bb->block_num, ins->inst_false_bb->block_num);
+               break;
+       case OP_LIVERANGE_START:
+       case OP_LIVERANGE_END:
+       case OP_GC_LIVENESS_DEF:
+       case OP_GC_LIVENESS_USE:
+               printf (" R%d", (int)ins->inst_c1);
+               break;
+       case OP_SEQ_POINT:
+               printf (" il: %x", (int)ins->inst_imm);
                break;
        default:
                break;
@@ -599,6 +705,12 @@ print_regtrack (RegTrack *t, int num)
        }
 }
 #else
+
+void
+mono_print_ji (const MonoJumpInfo *ji)
+{
+}
+
 void
 mono_print_ins_index (int i, MonoInst *ins)
 {
@@ -633,11 +745,23 @@ insert_after_ins (MonoBasicBlock *bb, MonoInst *ins, MonoInst **last, MonoInst*
        *last = to_insert;
 }
 
+static inline int
+get_vreg_bank (MonoCompile *cfg, int reg, int bank)
+{
+       if (vreg_is_ref (cfg, reg))
+               return MONO_REG_INT_REF;
+       else if (vreg_is_mp (cfg, reg))
+               return MONO_REG_INT_MP;
+       else
+               return bank;
+}
+
 /*
- * Force the spilling of the variable in the symbolic register 'reg'.
+ * Force the spilling of the variable in the symbolic register 'reg', and free 
+ * the hreg it was assigned to.
  */
-static int
-get_register_force_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, MonoInst *ins, int reg, int bank)
+static void
+spill_vreg (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, MonoInst *ins, int reg, int bank)
 {
        MonoInst *load;
        int i, sel, spill;
@@ -647,6 +771,9 @@ get_register_force_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **la
        symbolic = rs->symbolic [bank];
        sel = rs->vassign [reg];
 
+       /* the vreg we need to spill lives in another logical reg bank */
+       bank = translate_bank (cfg->rs, bank, sel);
+
        /*i = rs->isymbolic [sel];
        g_assert (i == reg);*/
        i = reg;
@@ -660,7 +787,7 @@ get_register_force_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **la
        MONO_INST_NEW (cfg, load, regbank_load_ops [bank]);
        load->dreg = sel;
        load->inst_basereg = cfg->frame_reg;
-       load->inst_offset = mono_spillvar_offset (cfg, spill, bank);
+       load->inst_offset = mono_spillvar_offset (cfg, spill, get_vreg_bank (cfg, reg, bank));
        insert_after_ins (bb, ins, last, load);
        DEBUG (printf ("SPILLED LOAD (%d at 0x%08lx(%%ebp)) R%d (freed %s)\n", spill, (long)load->inst_offset, i, mono_regname_full (sel, bank)));
        if (G_UNLIKELY (bank))
@@ -669,7 +796,10 @@ get_register_force_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **la
                i = mono_regstate_alloc_int (rs, regmask (sel));
        g_assert (i == sel);
 
-       return sel;
+       if (G_UNLIKELY (bank))
+               mono_regstate_free_general (rs, sel, bank);
+       else
+               mono_regstate_free_int (rs, sel);
 }
 
 /* This isn't defined on older glib versions and on some platforms */
@@ -681,7 +811,8 @@ static int
 get_register_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, MonoInst *ins, regmask_t regmask, int reg, int bank)
 {
        MonoInst *load;
-       int i, sel, spill;
+       int i, sel, spill, num_sregs;
+       int sregs [MONO_MAX_SRC_REGS];
        int *symbolic;
        MonoRegState *rs = cfg->rs;
 
@@ -689,21 +820,17 @@ get_register_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, Mo
 
        g_assert (bank < MONO_NUM_REGBANKS);
 
-       DEBUG (printf ("\tstart regmask to assign R%d: 0x%08" G_GUINT64_FORMAT " (R%d <- R%d R%d)\n", reg, (guint64)regmask, ins->dreg, ins->sreg1, ins->sreg2));
+       DEBUG (printf ("\tstart regmask to assign R%d: 0x%08" G_GUINT64_FORMAT " (R%d <- R%d R%d R%d)\n", reg, (guint64)regmask, ins->dreg, ins->sreg1, ins->sreg2, ins->sreg3));
        /* exclude the registers in the current instruction */
-       if ((sreg1_bank_ins (ins) == bank) && (reg != ins->sreg1) && (reg_is_freeable (ins->sreg1, bank) || (is_soft_reg (ins->sreg1, bank) && rs->vassign [ins->sreg1] >= 0))) {
-               if (is_soft_reg (ins->sreg1, bank))
-                       regmask &= ~ (regmask (rs->vassign [ins->sreg1]));
-               else
-                       regmask &= ~ (regmask (ins->sreg1));
-               DEBUG (printf ("\t\texcluding sreg1 %s\n", mono_regname_full (ins->sreg1, bank)));
-       }
-       if ((sreg2_bank_ins (ins) == bank) && (reg != ins->sreg2) && (reg_is_freeable (ins->sreg2, bank) || (is_soft_reg (ins->sreg2, bank) && rs->vassign [ins->sreg2] >= 0))) {
-               if (is_soft_reg (ins->sreg2, bank))
-                       regmask &= ~ (regmask (rs->vassign [ins->sreg2]));
-               else
-                       regmask &= ~ (regmask (ins->sreg2));
-               DEBUG (printf ("\t\texcluding sreg2 %s %d\n", mono_regname_full (ins->sreg2, bank), ins->sreg2));
+       num_sregs = mono_inst_get_src_registers (ins, sregs);
+       for (i = 0; i < num_sregs; ++i) {
+               if ((sreg_bank_ins (i, ins) == bank) && (reg != sregs [i]) && (reg_is_freeable (sregs [i], bank) || (is_soft_reg (sregs [i], bank) && rs->vassign [sregs [i]] >= 0))) {
+                       if (is_soft_reg (sregs [i], bank))
+                               regmask &= ~ (regmask (rs->vassign [sregs [i]]));
+                       else
+                               regmask &= ~ (regmask (sregs [i]));
+                       DEBUG (printf ("\t\texcluding sreg%d %s %d\n", i + 1, mono_regname_full (sregs [i], bank), sregs [i]));
+               }
        }
        if ((dreg_bank_ins (ins) == bank) && (reg != ins->dreg) && reg_is_freeable (ins->dreg, bank)) {
                regmask &= ~ (regmask (ins->dreg));
@@ -718,6 +845,10 @@ get_register_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, Mo
                for (i = 0; i < regbank_size [bank]; ++i) {
                        if (regmask & (regmask (i))) {
                                sel = i;
+
+                               /* the vreg we need to load lives in another logical bank */
+                               bank = translate_bank (cfg->rs, bank, sel);
+
                                DEBUG (printf ("\t\tselected register %s has assignment %d\n", mono_regname_full (sel, bank), rs->symbolic [bank] [sel]));
                                break;
                        }
@@ -747,7 +878,7 @@ get_register_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, Mo
        MONO_INST_NEW (cfg, load, regbank_load_ops [bank]);
        load->dreg = sel;
        load->inst_basereg = cfg->frame_reg;
-       load->inst_offset = mono_spillvar_offset (cfg, spill, bank);
+       load->inst_offset = mono_spillvar_offset (cfg, spill, get_vreg_bank (cfg, i, bank));
        insert_after_ins (bb, ins, last, load);
        DEBUG (printf ("\tSPILLED LOAD (%d at 0x%08lx(%%ebp)) R%d (freed %s)\n", spill, (long)load->inst_offset, i, mono_regname_full (sel, bank)));
        if (G_UNLIKELY (bank))
@@ -759,21 +890,25 @@ get_register_spilling (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, Mo
        return sel;
 }
 
+/*
+ * free_up_hreg:
+ *
+ *   Free up the hreg HREG by spilling the vreg allocated to it.
+ */
 static void
-free_up_reg (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, MonoInst *ins, int hreg, int bank)
+free_up_hreg (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, MonoInst *ins, int hreg, int bank)
 {
        if (G_UNLIKELY (bank)) {
-               if (!(cfg->rs->free_mask [1] & (regmask (hreg)))) {
+               if (!(cfg->rs->free_mask [bank] & (regmask (hreg)))) {
+                       bank = translate_bank (cfg->rs, bank, hreg);
                        DEBUG (printf ("\tforced spill of R%d\n", cfg->rs->symbolic [bank] [hreg]));
-                       get_register_force_spilling (cfg, bb, last, ins, cfg->rs->symbolic [bank] [hreg], bank);
-                       mono_regstate_free_general (cfg->rs, hreg, bank);
+                       spill_vreg (cfg, bb, last, ins, cfg->rs->symbolic [bank] [hreg], bank);
                }
        }
        else {
                if (!(cfg->rs->ifree_mask & (regmask (hreg)))) {
                        DEBUG (printf ("\tforced spill of R%d\n", cfg->rs->isymbolic [hreg]));
-                       get_register_force_spilling (cfg, bb, last, ins, cfg->rs->isymbolic [hreg], bank);
-                       mono_regstate_free_int (cfg->rs, hreg);
+                       spill_vreg (cfg, bb, last, ins, cfg->rs->isymbolic [hreg], bank);
                }
        }
 }
@@ -796,10 +931,24 @@ create_copy_ins (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst **last, int dest
        return copy;
 }
 
-static MonoInst*
-create_spilled_store (MonoCompile *cfg, MonoBasicBlock *bb, int spill, int reg, int prev_reg, MonoInst **last, MonoInst *ins, int bank)
+static inline const char*
+regbank_to_string (int bank)
+{
+       if (bank == MONO_REG_INT_REF)
+               return "REF ";
+       else if (bank == MONO_REG_INT_MP)
+               return "MP ";
+       else
+               return "";
+}
+
+static void
+create_spilled_store (MonoCompile *cfg, MonoBasicBlock *bb, int spill, int reg, int prev_reg, MonoInst **last, MonoInst *ins, MonoInst *insert_before, int bank)
 {
-       MonoInst *store;
+       MonoInst *store, *def;
+       
+       bank = get_vreg_bank (cfg, prev_reg, bank);
+
        MONO_INST_NEW (cfg, store, regbank_store_ops [bank]);
        store->sreg1 = reg;
        store->inst_destbasereg = cfg->frame_reg;
@@ -807,9 +956,20 @@ create_spilled_store (MonoCompile *cfg, MonoBasicBlock *bb, int spill, int reg,
        if (ins) {
                mono_bblock_insert_after_ins (bb, ins, store);
                *last = store;
+       } else if (insert_before) {
+               insert_before_ins (bb, insert_before, store);
+       } else {
+               g_assert_not_reached ();
+       }
+       DEBUG (printf ("\t%sSPILLED STORE (%d at 0x%08lx(%%ebp)) R%d (from %s)\n", regbank_to_string (bank), spill, (long)store->inst_offset, prev_reg, mono_regname_full (reg, bank)));
+
+       if (((bank == MONO_REG_INT_REF) || (bank == MONO_REG_INT_MP)) && cfg->compute_gc_maps) {
+               g_assert (prev_reg != -1);
+               MONO_INST_NEW (cfg, def, OP_GC_SPILL_SLOT_LIVENESS_DEF);
+               def->inst_c0 = spill;
+               def->inst_c1 = bank;
+               mono_bblock_insert_after_ins (bb, store, def);
        }
-       DEBUG (printf ("\tSPILLED STORE (%d at 0x%08lx(%%ebp)) R%d (from %s)\n", spill, (long)store->inst_offset, prev_reg, mono_regname_full (reg, bank)));
-       return store;
 }
 
 /* flags used in reginfo->flags */
@@ -865,6 +1025,8 @@ static inline void
 assign_reg (MonoCompile *cfg, MonoRegState *rs, int reg, int hreg, int bank)
 {
        if (G_UNLIKELY (bank)) {
+               int mirrored_bank;
+
                g_assert (reg >= regbank_size [bank]);
                g_assert (hreg < regbank_size [bank]);
                g_assert (! is_global_freg (hreg));
@@ -872,11 +1034,26 @@ assign_reg (MonoCompile *cfg, MonoRegState *rs, int reg, int hreg, int bank)
                rs->vassign [reg] = hreg;
                rs->symbolic [bank] [hreg] = reg;
                rs->free_mask [bank] &= ~ (regmask (hreg));
+
+               mirrored_bank = get_mirrored_bank (bank);
+               if (mirrored_bank == -1)
+                       return;
+
+               /* Make sure the other logical reg bank that this bank shares
+                * a single hard reg bank knows that this hard reg is not free.
+                */
+               rs->free_mask [mirrored_bank] = rs->free_mask [bank];
+
+               /* Mark the other logical bank that the this bank shares
+                * a single hard reg bank with as mirrored.
+                */
+               rs->symbolic [mirrored_bank] [hreg] = MONO_ARCH_BANK_MIRRORED;
+
        }
        else {
                g_assert (reg >= MONO_MAX_IREGS);
                g_assert (hreg < MONO_MAX_IREGS);
-#ifndef __arm__
+#ifndef TARGET_ARM
                /* this seems to trigger a gcc compilation bug sometime (hreg is 0) */
                g_assert (! is_global_ireg (hreg));
 #endif
@@ -898,6 +1075,8 @@ get_callee_mask (const char spec)
 static gint8 desc_to_fixed_reg [256];
 static gboolean desc_to_fixed_reg_inited = FALSE;
 
+#ifndef DISABLE_JIT
+
 /*
  * Local register allocation.
  * We first scan the list of instructions and we save the liveness info of
@@ -911,16 +1090,18 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
        MonoInst *ins, *prev, *last;
        MonoInst **tmp;
        MonoRegState *rs = cfg->rs;
-       int i, val, max;
+       int i, j, val, max;
        RegTrack *reginfo;
        const char *spec;
-       unsigned char spec_src1, spec_src2, spec_dest;
+       unsigned char spec_src1, spec_dest;
        int bank = 0;
 #if MONO_ARCH_USE_FPSTACK
        gboolean has_fp = FALSE;
        int fpstack [8];
        int sp = 0;
 #endif
+       int num_sregs = 0;
+       int sregs [MONO_MAX_SRC_REGS];
 
        if (!bb->code)
                return;
@@ -929,6 +1110,23 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                for (i = 0; i < 256; ++i)
                        desc_to_fixed_reg [i] = MONO_ARCH_INST_FIXED_REG (i);
                desc_to_fixed_reg_inited = TRUE;
+
+               /* Validate the cpu description against the info in mini-ops.h */
+#if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM)
+               for (i = OP_LOAD; i < OP_LAST; ++i) {
+                       const char *ispec;
+
+                       spec = ins_get_spec (i);
+                       ispec = INS_INFO (i);
+
+                       if ((spec [MONO_INST_DEST] && (ispec [MONO_INST_DEST] == ' ')))
+                               printf ("Instruction metadata for %s inconsistent.\n", mono_inst_name (i));
+                       if ((spec [MONO_INST_SRC1] && (ispec [MONO_INST_SRC1] == ' ')))
+                               printf ("Instruction metadata for %s inconsistent.\n", mono_inst_name (i));
+                       if ((spec [MONO_INST_SRC2] && (ispec [MONO_INST_SRC2] == ' ')))
+                               printf ("Instruction metadata for %s inconsistent.\n", mono_inst_name (i));
+               }
+#endif
        }
 
        rs->next_vreg = bb->max_vreg;
@@ -966,7 +1164,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 
                if ((ins->dreg != -1) && (ins->dreg < max)) {
                        memset (&reginfo [ins->dreg], 0, sizeof (RegTrack));
-#if SIZEOF_VOID_P == 4
+#if SIZEOF_REGISTER == 4
                        if (MONO_ARCH_INST_IS_REGPAIR (spec [MONO_INST_DEST])) {
                                /**
                                 * In the new IR, the two vregs of the regpair do not alias the
@@ -978,36 +1176,31 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                        }
 #endif
                }
-               if ((ins->sreg1 != -1) && (ins->sreg1 < max)) {
-                       memset (&reginfo [ins->sreg1], 0, sizeof (RegTrack));
-#if SIZEOF_VOID_P == 4
-                       if (MONO_ARCH_INST_IS_REGPAIR (spec [MONO_INST_SRC1])) {
-                               ins->sreg1 ++;
-                               memset (&reginfo [ins->sreg1 + 1], 0, sizeof (RegTrack));
-                       }
+
+               num_sregs = mono_inst_get_src_registers (ins, sregs);
+               for (j = 0; j < num_sregs; ++j) {
+                       g_assert (sregs [j] != -1);
+                       if (sregs [j] < max) {
+                               memset (&reginfo [sregs [j]], 0, sizeof (RegTrack));
+#if SIZEOF_REGISTER == 4
+                               if (MONO_ARCH_INST_IS_REGPAIR (spec [MONO_INST_SRC1 + j])) {
+                                       sregs [j]++;
+                                       memset (&reginfo [sregs [j] + 1], 0, sizeof (RegTrack));
+                               }
 #endif
-               }
-               if ((ins->sreg2 != -1) && (ins->sreg2 < max)) {
-                       memset (&reginfo [ins->sreg2], 0, sizeof (RegTrack));
-#if SIZEOF_VOID_P == 4
-                       if (MONO_ARCH_INST_IS_REGPAIR (spec [MONO_INST_SRC2])) {
-                               ins->sreg2 ++;
-                               memset (&reginfo [ins->sreg2 + 1], 0, sizeof (RegTrack));
                        }
-#endif
                }
+               mono_inst_set_src_registers (ins, sregs);
        }
 
        /*if (cfg->opt & MONO_OPT_COPYPROP)
                local_copy_prop (cfg, ins);*/
 
        i = 1;
-       DEBUG (printf ("\nLOCAL REGALLOC: BASIC BLOCK %d:\n", bb->block_num));
+       DEBUG (printf ("\nLOCAL REGALLOC BLOCK %d:\n", bb->block_num));
        /* forward pass on the instructions to collect register liveness info */
        MONO_BB_FOR_EACH_INS (bb, ins) {
                spec = ins_get_spec (ins->opcode);
-               spec_src1 = spec [MONO_INST_SRC1];
-               spec_src2 = spec [MONO_INST_SRC2];
                spec_dest = spec [MONO_INST_DEST];
 
                if (G_UNLIKELY (spec == MONO_ARCH_CPU_SPEC)) {
@@ -1016,51 +1209,45 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                
                DEBUG (mono_print_ins_index (i, ins));
 
+               num_sregs = mono_inst_get_src_registers (ins, sregs);
+
 #if MONO_ARCH_USE_FPSTACK
-               if (sreg1_is_fp (spec) || sreg2_is_fp (spec) || dreg_is_fp (spec))
+               if (dreg_is_fp (spec)) {
                        has_fp = TRUE;
-#endif
-
-               if (spec_src1) {
-                       bank = sreg1_bank (spec);
-                       g_assert (ins->sreg1 != -1);
-                       if (is_soft_reg (ins->sreg1, bank))
-                               /* This means the vreg is not local to this bb */
-                               g_assert (reginfo [ins->sreg1].born_in > 0);
-                       rs->vassign [ins->sreg1] = -1;
-                       //reginfo [ins->sreg1].prev_use = reginfo [ins->sreg1].last_use;
-                       //reginfo [ins->sreg1].last_use = i;
-                       if (MONO_ARCH_INST_IS_REGPAIR (spec_src2)) {
-                               /* The virtual register is allocated sequentially */
-                               rs->vassign [ins->sreg1 + 1] = -1;
-                               //reginfo [ins->sreg1 + 1].prev_use = reginfo [ins->sreg1 + 1].last_use;
-                               //reginfo [ins->sreg1 + 1].last_use = i;
-                               if (reginfo [ins->sreg1 + 1].born_in == 0 || reginfo [ins->sreg1 + 1].born_in > i)
-                                       reginfo [ins->sreg1 + 1].born_in = i;
-                       }
                } else {
-                       ins->sreg1 = -1;
+                       for (j = 0; j < num_sregs; ++j) {
+                               if (sreg_is_fp (j, spec))
+                                       has_fp = TRUE;
+                       }
                }
-               if (spec_src2) {
-                       bank = sreg2_bank (spec);
-                       g_assert (ins->sreg2 != -1);
-                       if (is_soft_reg (ins->sreg2, bank))
-                               /* This means the vreg is not local to this bb */
-                               g_assert (reginfo [ins->sreg2].born_in > 0);
-                       rs->vassign [ins->sreg2] = -1;
-                       //reginfo [ins->sreg2].prev_use = reginfo [ins->sreg2].last_use;
-                       //reginfo [ins->sreg2].last_use = i;
-                       if (MONO_ARCH_INST_IS_REGPAIR (spec_src2)) {
-                               /* The virtual register is allocated sequentially */
-                               rs->vassign [ins->sreg2 + 1] = -1;
-                               //reginfo [ins->sreg2 + 1].prev_use = reginfo [ins->sreg2 + 1].last_use;
-                               //reginfo [ins->sreg2 + 1].last_use = i;
-                               if (reginfo [ins->sreg2 + 1].born_in == 0 || reginfo [ins->sreg2 + 1].born_in > i)
-                                       reginfo [ins->sreg2 + 1].born_in = i;
+#endif
+
+               for (j = 0; j < num_sregs; ++j) {
+                       int sreg = sregs [j];
+                       int sreg_spec = spec [MONO_INST_SRC1 + j];
+                       if (sreg_spec) {
+                               bank = sreg_bank (j, spec);
+                               g_assert (sreg != -1);
+                               if (is_soft_reg (sreg, bank))
+                                       /* This means the vreg is not local to this bb */
+                                       g_assert (reginfo [sreg].born_in > 0);
+                               rs->vassign [sreg] = -1;
+                               //reginfo [ins->sreg2].prev_use = reginfo [ins->sreg2].last_use;
+                               //reginfo [ins->sreg2].last_use = i;
+                               if (MONO_ARCH_INST_IS_REGPAIR (sreg_spec)) {
+                                       /* The virtual register is allocated sequentially */
+                                       rs->vassign [sreg + 1] = -1;
+                                       //reginfo [ins->sreg2 + 1].prev_use = reginfo [ins->sreg2 + 1].last_use;
+                                       //reginfo [ins->sreg2 + 1].last_use = i;
+                                       if (reginfo [sreg + 1].born_in == 0 || reginfo [sreg + 1].born_in > i)
+                                               reginfo [sreg + 1].born_in = i;
+                               }
+                       } else {
+                               sregs [j] = -1;
                        }
-               } else {
-                       ins->sreg2 = -1;
                }
+               mono_inst_set_src_registers (ins, sregs);
+
                if (spec_dest) {
                        int dest_dreg;
 
@@ -1096,7 +1283,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                        ins->dreg = -1;
                }
 
-               if (spec [MONO_INST_CLOB] == 'c') {
+               if (spec [MONO_INST_CLOB] == 'c' && MONO_IS_CALL (ins)) {
                        /* A call instruction implicitly uses all registers in call->out_ireg_args */
 
                        MonoCallInst *call = (MonoCallInst*)ins;
@@ -1141,30 +1328,35 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 
        DEBUG (print_regtrack (reginfo, rs->next_vreg));
        MONO_BB_FOR_EACH_INS_REVERSE_SAFE (bb, prev, ins) {
-               int prev_dreg, prev_sreg1, prev_sreg2, clob_dreg;
-               int dest_dreg, dest_sreg1, dest_sreg2, clob_reg;
+               int prev_dreg, clob_dreg;
+               int dest_dreg, clob_reg;
+               int dest_sregs [MONO_MAX_SRC_REGS], prev_sregs [MONO_MAX_SRC_REGS];
                int dreg_high, sreg1_high;
-               regmask_t dreg_mask, sreg1_mask, sreg2_mask, mask;
-               regmask_t dreg_fixed_mask, sreg1_fixed_mask, sreg2_fixed_mask;
+               regmask_t dreg_mask, mask;
+               regmask_t sreg_masks [MONO_MAX_SRC_REGS], sreg_fixed_masks [MONO_MAX_SRC_REGS];
+               regmask_t dreg_fixed_mask;
                const unsigned char *ip;
                --i;
                spec = ins_get_spec (ins->opcode);
                spec_src1 = spec [MONO_INST_SRC1];
-               spec_src2 = spec [MONO_INST_SRC2];
                spec_dest = spec [MONO_INST_DEST];
                prev_dreg = -1;
-               prev_sreg2 = -1;
                clob_dreg = -1;
                clob_reg = -1;
                dest_dreg = -1;
-               dest_sreg1 = -1;
-               dest_sreg2 = -1;
-               prev_sreg1 = -1;
                dreg_high = -1;
                sreg1_high = -1;
                dreg_mask = get_callee_mask (spec_dest);
-               sreg1_mask = get_callee_mask (spec_src1);
-               sreg2_mask = get_callee_mask (spec_src2);
+               for (j = 0; j < MONO_MAX_SRC_REGS; ++j) {
+                       prev_sregs [j] = -1;
+                       sreg_masks [j] = get_callee_mask (spec [MONO_INST_SRC1 + j]);
+                       dest_sregs [j] = desc_to_fixed_reg [(int)spec [MONO_INST_SRC1 + j]];
+#ifdef MONO_ARCH_INST_FIXED_MASK
+                       sreg_fixed_masks [j] = MONO_ARCH_INST_FIXED_MASK (spec [MONO_INST_SRC1 + j]);
+#else
+                       sreg_fixed_masks [j] = 0;
+#endif
+               }
 
                DEBUG (printf ("processing:"));
                DEBUG (mono_print_ins_index (i, ins));
@@ -1176,57 +1368,123 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                /*
                 * FIXED REGS
                 */
-               dest_sreg1 = desc_to_fixed_reg [spec_src1];
-               dest_sreg2 = desc_to_fixed_reg [spec_src2];
                dest_dreg = desc_to_fixed_reg [spec_dest];
                clob_reg = desc_to_fixed_reg [(int)spec [MONO_INST_CLOB]];
-               sreg2_mask &= ~ (MONO_ARCH_INST_SREG2_MASK (spec));
+               sreg_masks [1] &= ~ (MONO_ARCH_INST_SREG2_MASK (spec));
 
 #ifdef MONO_ARCH_INST_FIXED_MASK
-               sreg1_fixed_mask = MONO_ARCH_INST_FIXED_MASK (spec_src1);
-               sreg2_fixed_mask = MONO_ARCH_INST_FIXED_MASK (spec_src2);
                dreg_fixed_mask = MONO_ARCH_INST_FIXED_MASK (spec_dest);
 #else
-               sreg1_fixed_mask = sreg2_fixed_mask = dreg_fixed_mask = 0;
+               dreg_fixed_mask = 0;
 #endif
 
+               num_sregs = mono_inst_get_src_registers (ins, sregs);
+
                /*
-                * TRACK FIXED SREG2
+                * TRACK FIXED SREG2, 3, ...
                 */
-               if (dest_sreg2 != -1) {
-                       if (rs->ifree_mask & (regmask (dest_sreg2))) {
-                               if (is_global_ireg (ins->sreg2)) {
+               for (j = 1; j < num_sregs; ++j) {
+                       int sreg = sregs [j];
+                       int dest_sreg = dest_sregs [j];
+
+                       if (dest_sreg == -1)
+                               continue;
+
+                       if (j == 2) {
+                               int k;
+
+                               /*
+                                * CAS.
+                                * We need to special case this, since on x86, there are only 3
+                                * free registers, and the code below assigns one of them to
+                                * sreg, so we can run out of registers when trying to assign
+                                * dreg. Instead, we just set up the register masks, and let the
+                                * normal sreg2 assignment code handle this. It would be nice to
+                                * do this for all the fixed reg cases too, but there is too much
+                                * risk of breakage.
+                                */
+
+                               /* Make sure sreg will be assigned to dest_sreg, and the other sregs won't */
+                               sreg_masks [j] = regmask (dest_sreg);
+                               for (k = 0; k < num_sregs; ++k) {
+                                       if (k != j)
+                                               sreg_masks [k] &= ~ (regmask (dest_sreg));
+                               }                                               
+
+                               /*
+                                * Spill sreg1/2 if they are assigned to dest_sreg.
+                                */
+                               for (k = 0; k < num_sregs; ++k) {
+                                       if (k != j && is_soft_reg (sregs [k], 0) && rs->vassign [sregs [k]] == dest_sreg)
+                                               free_up_hreg (cfg, bb, tmp, ins, dest_sreg, 0);
+                               }
+
+                               /*
+                                * We can also run out of registers while processing sreg2 if sreg3 is
+                                * assigned to another hreg, so spill sreg3 now.
+                                */
+                               if (is_soft_reg (sreg, 0) && rs->vassign [sreg] >= 0 && rs->vassign [sreg] != dest_sreg) {
+                                       spill_vreg (cfg, bb, tmp, ins, sreg, 0);
+                               }
+                               continue;
+                       }
+
+                       if (rs->ifree_mask & (regmask (dest_sreg))) {
+                               if (is_global_ireg (sreg)) {
+                                       int k;
                                        /* Argument already in hard reg, need to copy */
-                                       MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg2, ins->sreg2, NULL, ip, 0);
+                                       MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg, sreg, NULL, ip, 0);
                                        insert_before_ins (bb, ins, copy);
-                               }
-                               else {
-                                       val = rs->vassign [ins->sreg2];
+                                       for (k = 0; k < num_sregs; ++k) {
+                                               if (k != j)
+                                                       sreg_masks [k] &= ~ (regmask (dest_sreg));
+                                       }
+                               } else {
+                                       val = rs->vassign [sreg];
                                        if (val == -1) {
-                                               DEBUG (printf ("\tshortcut assignment of R%d to %s\n", ins->sreg2, mono_arch_regname (dest_sreg2)));
-                                               assign_reg (cfg, rs, ins->sreg2, dest_sreg2, 0);
+                                               DEBUG (printf ("\tshortcut assignment of R%d to %s\n", sreg, mono_arch_regname (dest_sreg)));
+                                               assign_reg (cfg, rs, sreg, dest_sreg, 0);
                                        } else if (val < -1) {
                                                /* FIXME: */
                                                g_assert_not_reached ();
                                        } else {
                                                /* Argument already in hard reg, need to copy */
-                                               MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg2, val, NULL, ip, 0);
+                                               MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg, val, NULL, ip, 0);
+                                               int k;
+
                                                insert_before_ins (bb, ins, copy);
+                                               for (k = 0; k < num_sregs; ++k) {
+                                                       if (k != j)
+                                                               sreg_masks [k] &= ~ (regmask (dest_sreg));
+                                               }
+                                               /* 
+                                                * Prevent the dreg from being allocate to dest_sreg 
+                                                * too, since it could force sreg1 to be allocated to 
+                                                * the same reg on x86.
+                                                */
+                                               dreg_mask &= ~ (regmask (dest_sreg));
                                        }
                                }
                        } else {
                                gboolean need_spill = TRUE;
                                gboolean need_assign = TRUE;
+                               int k;
 
-                               dreg_mask &= ~ (regmask (dest_sreg2));
-                               sreg1_mask &= ~ (regmask (dest_sreg2));
+                               dreg_mask &= ~ (regmask (dest_sreg));
+                               for (k = 0; k < num_sregs; ++k) {
+                                       if (k != j)
+                                               sreg_masks [k] &= ~ (regmask (dest_sreg));
+                               }
 
                                /* 
                                 * First check if dreg is assigned to dest_sreg2, since we
                                 * can't spill a dreg.
                                 */
-                               val = rs->vassign [ins->dreg];
-                               if (val == dest_sreg2 && ins->dreg != ins->sreg2) {
+                               if (spec [MONO_INST_DEST])
+                                       val = rs->vassign [ins->dreg];
+                               else
+                                       val = -1;
+                               if (val == dest_sreg && ins->dreg != sreg) {
                                        /* 
                                         * the destination register is already assigned to 
                                         * dest_sreg2: we need to allocate another register for it 
@@ -1235,24 +1493,24 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                        int new_dest;
                                        new_dest = alloc_int_reg (cfg, bb, tmp, ins, dreg_mask, ins->dreg, &reginfo [ins->dreg]);
                                        g_assert (new_dest >= 0);
-                                       DEBUG (printf ("\tchanging dreg R%d to %s from %s\n", ins->dreg, mono_arch_regname (new_dest), mono_arch_regname (dest_sreg2)));
+                                       DEBUG (printf ("\tchanging dreg R%d to %s from %s\n", ins->dreg, mono_arch_regname (new_dest), mono_arch_regname (dest_sreg)));
 
                                        prev_dreg = ins->dreg;
                                        assign_reg (cfg, rs, ins->dreg, new_dest, 0);
                                        clob_dreg = ins->dreg;
-                                       create_copy_ins (cfg, bb, tmp, dest_sreg2, new_dest, ins, ip, 0);
-                                       mono_regstate_free_int (rs, dest_sreg2);
+                                       create_copy_ins (cfg, bb, tmp, dest_sreg, new_dest, ins, ip, 0);
+                                       mono_regstate_free_int (rs, dest_sreg);
                                        need_spill = FALSE;
                                }
 
-                               if (is_global_ireg (ins->sreg2)) {
-                                       MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg2, ins->sreg2, NULL, ip, 0);
+                               if (is_global_ireg (sreg)) {
+                                       MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg, sreg, NULL, ip, 0);
                                        insert_before_ins (bb, ins, copy);
                                        need_assign = FALSE;
                                }
                                else {
-                                       val = rs->vassign [ins->sreg2];
-                                       if (val == dest_sreg2) {
+                                       val = rs->vassign [sreg];
+                                       if (val == dest_sreg) {
                                                /* sreg2 is already assigned to the correct register */
                                                need_spill = FALSE;
                                        } else if (val < -1) {
@@ -1265,32 +1523,29 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                                 * instruction. So we spill sreg2 so it can be allocated to
                                                 * dest_sreg2.
                                                 */
-                                               DEBUG (printf ("\tforced spill of R%d\n", ins->sreg2));
-                                               free_up_reg (cfg, bb, tmp, ins, val, 0);
+                                               free_up_hreg (cfg, bb, tmp, ins, val, 0);
                                        }
                                }
 
                                if (need_spill) {
-                                       DEBUG (printf ("\tforced spill of R%d\n", rs->isymbolic [dest_sreg2]));
-                                       free_up_reg (cfg, bb, tmp, ins, dest_sreg2, 0);
+                                       free_up_hreg (cfg, bb, tmp, ins, dest_sreg, 0);
                                }
 
                                if (need_assign) {
-                                       if (rs->vassign [ins->sreg2] < -1) {
-                                               MonoInst *store;
+                                       if (rs->vassign [sreg] < -1) {
                                                int spill;
 
                                                /* Need to emit a spill store */
-                                               spill = - rs->vassign [ins->sreg2] - 1;
-                                               store = create_spilled_store (cfg, bb, spill, dest_sreg2, ins->sreg2, tmp, NULL, bank);                                         
-                                               insert_before_ins (bb, ins, store);
+                                               spill = - rs->vassign [sreg] - 1;
+                                               create_spilled_store (cfg, bb, spill, dest_sreg, sreg, tmp, NULL, ins, bank);
                                        }
                                        /* force-set sreg2 */
-                                       assign_reg (cfg, rs, ins->sreg2, dest_sreg2, 0);
+                                       assign_reg (cfg, rs, sregs [j], dest_sreg, 0);
                                }
                        }
-                       ins->sreg2 = dest_sreg2;
+                       sregs [j] = dest_sreg;
                }
+               mono_inst_set_src_registers (ins, sregs);
 
                /*
                 * TRACK DREG
@@ -1305,16 +1560,17 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                         * The dest reg is read by the instruction, not written, so
                         * avoid allocating sreg1/sreg2 to the same reg.
                         */
-                       if (!dest_sreg1 != -1)
-                               dreg_mask &= ~ (regmask (dest_sreg1));
-                       if (dest_sreg2 != -1)
-                               dreg_mask &= ~ (regmask (dest_sreg2));
+                       if (dest_sregs [0] != -1)
+                               dreg_mask &= ~ (regmask (dest_sregs [0]));
+                       for (j = 1; j < num_sregs; ++j) {
+                               if (dest_sregs [j] != -1)
+                                       dreg_mask &= ~ (regmask (dest_sregs [j]));
+                       }
 
                        val = rs->vassign [ins->dreg];
                        if (is_soft_reg (ins->dreg, bank) && (val >= 0) && (!(regmask (val) & dreg_mask))) {
                                /* DREG is already allocated to a register needed for sreg1 */
-                               get_register_force_spilling (cfg, bb, tmp, ins, ins->dreg, 0);
-                               mono_regstate_free_int (rs, val);
+                           spill_vreg (cfg, bb, tmp, ins, ins->dreg, 0);
                        }
                }
 
@@ -1329,13 +1585,13 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 
                        if (dest_dreg != -1) {
                                if (rs->vassign [ins->dreg] != dest_dreg)
-                                       free_up_reg (cfg, bb, tmp, ins, dest_dreg, 0);
+                                       free_up_hreg (cfg, bb, tmp, ins, dest_dreg, 0);
 
                                dreg2 = ins->dreg + 1;
                                dest_dreg2 = MONO_ARCH_INST_REGPAIR_REG2 (spec_dest, dest_dreg);
                                if (dest_dreg2 != -1) {
                                        if (rs->vassign [dreg2] != dest_dreg2)
-                                               free_up_reg (cfg, bb, tmp, ins, dest_dreg2, 0);
+                                               free_up_hreg (cfg, bb, tmp, ins, dest_dreg2, 0);
                                }
                        }
                }
@@ -1371,7 +1627,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                val = alloc_reg (cfg, bb, tmp, ins, dreg_mask, ins->dreg, &reginfo [ins->dreg], bank);
                                assign_reg (cfg, rs, ins->dreg, val, bank);
                                if (spill)
-                                       create_spilled_store (cfg, bb, spill, val, prev_dreg, tmp, ins, bank);
+                                       create_spilled_store (cfg, bb, spill, val, prev_dreg, tmp, ins, NULL, bank);
                        }
 
                        DEBUG (printf ("\tassigned dreg %s to dest R%d\n", mono_regname_full (val, bank), ins->dreg));
@@ -1386,7 +1642,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                        g_assert (prev_dreg > -1);
                        g_assert (!is_global_ireg (rs->vassign [prev_dreg]));
                        mask = regpair_reg2_mask (spec_dest, rs->vassign [prev_dreg]);
-#ifdef __i386__
+#ifdef TARGET_X86
                        /* bug #80489 */
                        mask &= ~regmask (X86_ECX);
 #endif
@@ -1401,7 +1657,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                if (val < 0)
                                        val = get_register_spilling (cfg, bb, tmp, ins, mask, reg2, bank);
                                if (spill)
-                                       create_spilled_store (cfg, bb, spill, val, reg2, tmp, ins, bank);
+                                       create_spilled_store (cfg, bb, spill, val, reg2, tmp, ins, NULL, bank);
                        }
                        else {
                                if (! (mask & (regmask (val)))) {
@@ -1450,12 +1706,16 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                        ins->dreg = dest_dreg;
 
                        if (G_UNLIKELY (bank)) {
-                               if (rs->symbolic [bank] [dest_dreg] >= regbank_size [bank])
-                                       free_up_reg (cfg, bb, tmp, ins, dest_dreg, bank);
+                               /* the register we need to free up may be used in another logical regbank
+                                * so do a translate just in case.
+                                */
+                               int translated_bank = translate_bank (cfg->rs, bank, dest_dreg);
+                               if (rs->symbolic [translated_bank] [dest_dreg] >= regbank_size [translated_bank])
+                                       free_up_hreg (cfg, bb, tmp, ins, dest_dreg, translated_bank);
                        }
                        else {
                                if (rs->isymbolic [dest_dreg] >= MONO_MAX_IREGS)
-                                       free_up_reg (cfg, bb, tmp, ins, dest_dreg, bank);
+                                       free_up_hreg (cfg, bb, tmp, ins, dest_dreg, bank);
                        }
                }
 
@@ -1464,10 +1724,9 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                         * The dest reg is read by the instruction, not written, so
                         * avoid allocating sreg1/sreg2 to the same reg.
                         */
-                       if (!sreg1_bank (spec))
-                               sreg1_mask &= ~ (regmask (ins->dreg));
-                       if (!sreg2_bank (spec))
-                               sreg2_mask &= ~ (regmask (ins->dreg));
+                       for (j = 0; j < num_sregs; ++j)
+                               if (!sreg_bank (j, spec))
+                                       sreg_masks [j] &= ~ (regmask (ins->dreg));
                }
 
                /*
@@ -1475,8 +1734,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                 */
                if ((clob_reg != -1) && (!(rs->ifree_mask & (regmask (clob_reg))))) {
                        DEBUG (printf ("\tforced spill of clobbered reg R%d\n", rs->isymbolic [clob_reg]));
-                       get_register_force_spilling (cfg, bb, tmp, ins, rs->isymbolic [clob_reg], 0);
-                       mono_regstate_free_int (rs, clob_reg);
+                       free_up_hreg (cfg, bb, tmp, ins, clob_reg, 0);
                }
 
                if (spec [MONO_INST_CLOB] == 'c') {
@@ -1504,7 +1762,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                        s = regmask (j);
                                        if ((clob_mask & s) && !(rs->ifree_mask & s) && (j != ins->sreg1)) {
                                                if ((j != dreg) && (j != dreg2))
-                                                       get_register_force_spilling (cfg, bb, tmp, ins, rs->isymbolic [j], 0);
+                                                       free_up_hreg (cfg, bb, tmp, ins, j, 0);
                                                else if (rs->isymbolic [j])
                                                        /* The hreg is assigned to the dreg of this instruction */
                                                        rs->vassign [rs->isymbolic [j]] = -1;
@@ -1522,10 +1780,18 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                                dreg = -1;
 
                                        for (j = 0; j < regbank_size [cur_bank]; ++j) {
+
+                                               /* we are looping though the banks in the outer loop
+                                                * so, we don't need to deal with mirrored hregs
+                                                * because we will get them in one of the other bank passes.
+                                                */
+                                               if (is_hreg_mirrored (rs, cur_bank, j))
+                                                       continue;
+
                                                s = regmask (j);
                                                if ((clob_mask & s) && !(rs->free_mask [cur_bank] & s) && (j != ins->sreg1)) {
                                                        if (j != dreg)
-                                                               get_register_force_spilling (cfg, bb, tmp, ins, rs->symbolic [cur_bank] [j], cur_bank);
+                                                               free_up_hreg (cfg, bb, tmp, ins, j, cur_bank);
                                                        else if (rs->symbolic [cur_bank] [j])
                                                                /* The hreg is assigned to the dreg of this instruction */
                                                                rs->vassign [rs->symbolic [cur_bank] [j]] = -1;
@@ -1539,7 +1805,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                /*
                 * TRACK ARGUMENT REGS
                 */
-               if (spec [MONO_INST_CLOB] == 'c') {
+               if (spec [MONO_INST_CLOB] == 'c' && MONO_IS_CALL (ins)) {
                        MonoCallInst *call = (MonoCallInst*)ins;
                        GSList *list;
 
@@ -1565,7 +1831,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 
                                        assign_reg (cfg, rs, reg, hreg, 0);
 
-                                       sreg1_mask &= ~(regmask (hreg));
+                                       sreg_masks [0] &= ~(regmask (hreg));
 
                                        DEBUG (printf ("\tassigned arg reg %s to R%d\n", mono_arch_regname (hreg), reg));
 
@@ -1597,7 +1863,10 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                 */
                bank = sreg1_bank (spec);
                if (MONO_ARCH_INST_IS_REGPAIR (spec_dest) && (spec [MONO_INST_CLOB] == '1')) {
-                       g_assert (is_soft_reg (ins->sreg1, bank));
+                       int sreg1 = sregs [0];
+                       int dest_sreg1 = dest_sregs [0];
+
+                       g_assert (is_soft_reg (sreg1, bank));
 
                        /* To simplify things, we allocate the same regpair to sreg1 and dreg */
                        if (dest_sreg1 != -1)
@@ -1605,68 +1874,66 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                        val = mono_regstate_alloc_int (rs, regmask (ins->dreg));
                        g_assert (val >= 0);
 
-                       if (rs->vassign [ins->sreg1] >= 0 && rs->vassign [ins->sreg1] != val)
+                       if (rs->vassign [sreg1] >= 0 && rs->vassign [sreg1] != val)
                                // FIXME:
                                g_assert_not_reached ();
 
-                       assign_reg (cfg, rs, ins->sreg1, val, bank);
+                       assign_reg (cfg, rs, sreg1, val, bank);
 
-                       DEBUG (printf ("\tassigned sreg1-low %s to R%d\n", mono_regname_full (val, bank), ins->sreg1));
+                       DEBUG (printf ("\tassigned sreg1-low %s to R%d\n", mono_regname_full (val, bank), sreg1));
 
                        g_assert ((regmask (dreg_high)) & regpair_reg2_mask (spec_src1, ins->dreg));
                        val = mono_regstate_alloc_int (rs, regmask (dreg_high));
                        g_assert (val >= 0);
 
-                       if (rs->vassign [ins->sreg1 + 1] >= 0 && rs->vassign [ins->sreg1 + 1] != val)
+                       if (rs->vassign [sreg1 + 1] >= 0 && rs->vassign [sreg1 + 1] != val)
                                // FIXME:
                                g_assert_not_reached ();
 
-                       assign_reg (cfg, rs, ins->sreg1 + 1, val, bank);
+                       assign_reg (cfg, rs, sreg1 + 1, val, bank);
 
-                       DEBUG (printf ("\tassigned sreg1-high %s to R%d\n", mono_regname_full (val, bank), ins->sreg1 + 1));
+                       DEBUG (printf ("\tassigned sreg1-high %s to R%d\n", mono_regname_full (val, bank), sreg1 + 1));
 
                        /* Skip rest of this section */
-                       dest_sreg1 = -1;
+                       dest_sregs [0] = -1;
                }
 
-               if (sreg1_fixed_mask) {
+               if (sreg_fixed_masks [0]) {
                        g_assert (!bank);
-                       if (is_global_ireg (ins->sreg1)) {
+                       if (is_global_ireg (sregs [0])) {
                                /* 
                                 * The argument is already in a hard reg, but that reg is
                                 * not usable by this instruction, so allocate a new one.
                                 */
-                               val = mono_regstate_alloc_int (rs, sreg1_fixed_mask);
+                               val = mono_regstate_alloc_int (rs, sreg_fixed_masks [0]);
                                if (val < 0)
-                                       val = get_register_spilling (cfg, bb, tmp, ins, sreg1_fixed_mask, -1, bank);
+                                       val = get_register_spilling (cfg, bb, tmp, ins, sreg_fixed_masks [0], -1, bank);
                                mono_regstate_free_int (rs, val);
-                               dest_sreg1 = val;
+                               dest_sregs [0] = val;
 
                                /* Fall through to the dest_sreg1 != -1 case */
                        }
                        else
-                               sreg1_mask &= sreg1_fixed_mask;
+                               sreg_masks [0] &= sreg_fixed_masks [0];
                }
 
-               if (dest_sreg1 != -1) {
-                       sreg1_mask = regmask (dest_sreg1);
+               if (dest_sregs [0] != -1) {
+                       sreg_masks [0] = regmask (dest_sregs [0]);
 
-                       if ((rs->vassign [ins->sreg1] != dest_sreg1) && !(rs->ifree_mask & (regmask (dest_sreg1)))) {
-                               DEBUG (printf ("\tforced spill of R%d\n", rs->isymbolic [dest_sreg1]));
-                               get_register_force_spilling (cfg, bb, tmp, ins, rs->isymbolic [dest_sreg1], 0);
-                               mono_regstate_free_int (rs, dest_sreg1);
+                       if ((rs->vassign [sregs [0]] != dest_sregs [0]) && !(rs->ifree_mask & (regmask (dest_sregs [0])))) {
+                               free_up_hreg (cfg, bb, tmp, ins, dest_sregs [0], 0);
                        }
-                       if (is_global_ireg (ins->sreg1)) {
+                       if (is_global_ireg (sregs [0])) {
                                /* The argument is already in a hard reg, need to copy */
-                               MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg1, ins->sreg1, NULL, ip, 0);
+                               MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sregs [0], sregs [0], NULL, ip, 0);
                                insert_before_ins (bb, ins, copy);
-                               ins->sreg1 = dest_sreg1;
+                               sregs [0] = dest_sregs [0];
                        }
                }
 
-               if (is_soft_reg (ins->sreg1, bank)) {
-                       val = rs->vassign [ins->sreg1];
-                       prev_sreg1 = ins->sreg1;
+               if (is_soft_reg (sregs [0], bank)) {
+                       val = rs->vassign [sregs [0]];
+                       prev_sregs [0] = sregs [0];
                        if (val < 0) {
                                int spill = 0;
                                if (val < -1) {
@@ -1679,48 +1946,51 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                         * Allocate the same hreg to sreg1 as well so the 
                                         * peephole can get rid of the move.
                                         */
-                                       sreg1_mask = regmask (ins->dreg);
+                                       sreg_masks [0] = regmask (ins->dreg);
                                }
 
                                if (spec [MONO_INST_CLOB] == '1' && !dreg_bank (spec) && (rs->ifree_mask & (regmask (ins->dreg))))
                                        /* Allocate the same reg to sreg1 to avoid a copy later */
-                                       sreg1_mask = regmask (ins->dreg);
+                                       sreg_masks [0] = regmask (ins->dreg);
 
-                               val = alloc_reg (cfg, bb, tmp, ins, sreg1_mask, ins->sreg1, &reginfo [ins->sreg1], bank);
-                               assign_reg (cfg, rs, ins->sreg1, val, bank);
-                               DEBUG (printf ("\tassigned sreg1 %s to R%d\n", mono_regname_full (val, bank), ins->sreg1));
+                               val = alloc_reg (cfg, bb, tmp, ins, sreg_masks [0], sregs [0], &reginfo [sregs [0]], bank);
+                               assign_reg (cfg, rs, sregs [0], val, bank);
+                               DEBUG (printf ("\tassigned sreg1 %s to R%d\n", mono_regname_full (val, bank), sregs [0]));
 
                                if (spill) {
-                                       MonoInst *store = create_spilled_store (cfg, bb, spill, val, prev_sreg1, tmp, NULL, bank);
                                        /*
                                         * Need to insert before the instruction since it can
                                         * overwrite sreg1.
                                         */
-                                       insert_before_ins (bb, ins, store);
+                                       create_spilled_store (cfg, bb, spill, val, prev_sregs [0], tmp, NULL, ins, bank);
                                }
                        }
-                       else if ((dest_sreg1 != -1) && (dest_sreg1 != val)) {
-                               MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sreg1, val, NULL, ip, bank);
+                       else if ((dest_sregs [0] != -1) && (dest_sregs [0] != val)) {
+                               MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sregs [0], val, NULL, ip, bank);
                                insert_before_ins (bb, ins, copy);
-                               sreg2_mask &= ~(regmask (dest_sreg1));
-                               val = dest_sreg1;
+                               for (j = 1; j < num_sregs; ++j)
+                                       sreg_masks [j] &= ~(regmask (dest_sregs [0]));
+                               val = dest_sregs [0];
                        }
                                
-                       ins->sreg1 = val;
+                       sregs [0] = val;
                }
                else {
-                       prev_sreg1 = -1;
+                       prev_sregs [0] = -1;
                }
-               sreg2_mask &= ~(regmask (ins->sreg1));
+               mono_inst_set_src_registers (ins, sregs);
+
+               for (j = 1; j < num_sregs; ++j)
+                       sreg_masks [j] &= ~(regmask (sregs [0]));
 
                /* Handle the case when sreg1 is a regpair but dreg is not */
                if (MONO_ARCH_INST_IS_REGPAIR (spec_src1) && (spec [MONO_INST_CLOB] != '1')) {
-                       int reg2 = prev_sreg1 + 1;
+                       int reg2 = prev_sregs [0] + 1;
 
                        g_assert (!bank);
-                       g_assert (prev_sreg1 > -1);
-                       g_assert (!is_global_ireg (rs->vassign [prev_sreg1]));
-                       mask = regpair_reg2_mask (spec_src1, rs->vassign [prev_sreg1]);
+                       g_assert (prev_sregs [0] > -1);
+                       g_assert (!is_global_ireg (rs->vassign [prev_sregs [0]]));
+                       mask = regpair_reg2_mask (spec_src1, rs->vassign [prev_sregs [0]]);
                        val = rs->vassign [reg2];
                        if (val < 0) {
                                int spill = 0;
@@ -1758,21 +2028,21 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                }
 
                /* Handle dreg==sreg1 */
-               if (((dreg_is_fp (spec) && sreg1_is_fp (spec)) || spec [MONO_INST_CLOB] == '1') && ins->dreg != ins->sreg1) {
+               if (((dreg_is_fp (spec) && sreg1_is_fp (spec)) || spec [MONO_INST_CLOB] == '1') && ins->dreg != sregs [0]) {
                        MonoInst *sreg2_copy = NULL;
                        MonoInst *copy;
                        int bank = reg_bank (spec_src1);
 
-                       if (ins->dreg == ins->sreg2) {
+                       if (ins->dreg == sregs [1]) {
                                /* 
                                 * copying sreg1 to dreg could clobber sreg2, so allocate a new
                                 * register for it.
                                 */
-                               int reg2 = alloc_reg (cfg, bb, tmp, ins, dreg_mask, ins->sreg2, NULL, bank);
+                               int reg2 = alloc_reg (cfg, bb, tmp, ins, dreg_mask, sregs [1], NULL, bank);
 
-                               DEBUG (printf ("\tneed to copy sreg2 %s to reg %s\n", mono_regname_full (ins->sreg2, bank), mono_regname_full (reg2, bank)));
-                               sreg2_copy = create_copy_ins (cfg, bb, tmp, reg2, ins->sreg2, NULL, ip, bank);
-                               prev_sreg2 = ins->sreg2 = reg2;
+                               DEBUG (printf ("\tneed to copy sreg2 %s to reg %s\n", mono_regname_full (sregs [1], bank), mono_regname_full (reg2, bank)));
+                               sreg2_copy = create_copy_ins (cfg, bb, tmp, reg2, sregs [1], NULL, ip, bank);
+                               prev_sregs [1] = sregs [1] = reg2;
 
                                if (G_UNLIKELY (bank))
                                        mono_regstate_free_general (rs, reg2, bank);
@@ -1782,7 +2052,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 
                        if (MONO_ARCH_INST_IS_REGPAIR (spec_src1)) {
                                /* Copying sreg1_high to dreg could also clobber sreg2 */
-                               if (rs->vassign [prev_sreg1 + 1] == ins->sreg2)
+                               if (rs->vassign [prev_sregs [0] + 1] == sregs [1])
                                        /* FIXME: */
                                        g_assert_not_reached ();
 
@@ -1790,12 +2060,12 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                 * sreg1 and dest are already allocated to the same regpair by the
                                 * SREG1 allocation code.
                                 */
-                               g_assert (ins->sreg1 == ins->dreg);
+                               g_assert (sregs [0] == ins->dreg);
                                g_assert (dreg_high == sreg1_high);
                        }
 
-                       DEBUG (printf ("\tneed to copy sreg1 %s to dreg %s\n", mono_regname_full (ins->sreg1, bank), mono_regname_full (ins->dreg, bank)));
-                       copy = create_copy_ins (cfg, bb, tmp, ins->dreg, ins->sreg1, NULL, ip, bank);
+                       DEBUG (printf ("\tneed to copy sreg1 %s to dreg %s\n", mono_regname_full (sregs [0], bank), mono_regname_full (ins->dreg, bank)));
+                       copy = create_copy_ins (cfg, bb, tmp, ins->dreg, sregs [0], NULL, ip, bank);
                        insert_before_ins (bb, ins, copy);
 
                        if (sreg2_copy)
@@ -1805,43 +2075,88 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                         * Need to prevent sreg2 to be allocated to sreg1, since that
                         * would screw up the previous copy.
                         */
-                       sreg2_mask &= ~ (regmask (ins->sreg1));
+                       sreg_masks [1] &= ~ (regmask (sregs [0]));
                        /* we set sreg1 to dest as well */
-                       prev_sreg1 = ins->sreg1 = ins->dreg;
-                       sreg2_mask &= ~ (regmask (ins->dreg));
+                       prev_sregs [0] = sregs [0] = ins->dreg;
+                       sreg_masks [1] &= ~ (regmask (ins->dreg));
                }
+               mono_inst_set_src_registers (ins, sregs);
 
                /*
-                * TRACK SREG2
+                * TRACK SREG2, 3, ...
                 */
-               bank = sreg2_bank (spec);
-               if (MONO_ARCH_INST_IS_REGPAIR (spec_src2))
-                       g_assert_not_reached ();
-               if (is_soft_reg (ins->sreg2, bank)) {
-                       val = rs->vassign [ins->sreg2];
+               for (j = 1; j < num_sregs; ++j) {
+                       int k;
 
-                       if (val < 0) {
-                               int spill = 0;
-                               if (val < -1) {
-                                       /* the register gets spilled after this inst */
-                                       spill = -val -1;
-                               }
-                               val = alloc_reg (cfg, bb, tmp, ins, sreg2_mask, ins->sreg2, &reginfo [ins->sreg2], bank);
-                               assign_reg (cfg, rs, ins->sreg2, val, bank);
-                               DEBUG (printf ("\tassigned sreg2 %s to R%d\n", mono_regname_full (val, bank), ins->sreg2));
-                               if (spill) {
-                                       MonoInst *store = create_spilled_store (cfg, bb, spill, val, prev_sreg2, tmp, NULL, bank);
+                       bank = sreg_bank (j, spec);
+                       if (MONO_ARCH_INST_IS_REGPAIR (spec [MONO_INST_SRC1 + j]))
+                               g_assert_not_reached ();
+
+                       if (dest_sregs [j] != -1 && is_global_ireg (sregs [j])) {
+                               /*
+                                * Argument already in a global hard reg, copy it to the fixed reg, without
+                                * allocating it to the fixed reg.
+                                */
+                               MonoInst *copy = create_copy_ins (cfg, bb, tmp, dest_sregs [j], sregs [j], NULL, ip, 0);
+                               insert_before_ins (bb, ins, copy);
+                               sregs [j] = dest_sregs [j];
+                       } else if (is_soft_reg (sregs [j], bank)) {
+                               val = rs->vassign [sregs [j]];
+
+                               if (dest_sregs [j] != -1 && val >= 0 && dest_sregs [j] != val) {
                                        /*
-                                        * Need to insert before the instruction since it can
-                                        * overwrite sreg2.
+                                        * The sreg is already allocated to a hreg, but not to the fixed
+                                        * reg required by the instruction. Spill the sreg, so it can be
+                                        * allocated to the fixed reg by the code below.
                                         */
-                                       insert_before_ins (bb, ins, store);
+                                       /* Currently, this code should only be hit for CAS */
+                                       spill_vreg (cfg, bb, tmp, ins, sregs [j], 0);
+                                       val = rs->vassign [sregs [j]];
                                }
+
+                               if (val < 0) {
+                                       int spill = 0;
+                                       if (val < -1) {
+                                               /* the register gets spilled after this inst */
+                                               spill = -val -1;
+                                       }
+                                       val = alloc_reg (cfg, bb, tmp, ins, sreg_masks [j], sregs [j], &reginfo [sregs [j]], bank);
+                                       assign_reg (cfg, rs, sregs [j], val, bank);
+                                       DEBUG (printf ("\tassigned sreg%d %s to R%d\n", j + 1, mono_regname_full (val, bank), sregs [j]));
+                                       if (spill) {
+                                               /*
+                                                * Need to insert before the instruction since it can
+                                                * overwrite sreg2.
+                                                */
+                                               create_spilled_store (cfg, bb, spill, val, sregs [j], tmp, NULL, ins, bank);
+                                       }
+                               }
+                               sregs [j] = val;
+                               for (k = j + 1; k < num_sregs; ++k)
+                                       sreg_masks [k] &= ~ (regmask (sregs [j]));
+                       }
+                       else {
+                               prev_sregs [j] = -1;
                        }
-                       ins->sreg2 = val;
                }
-               else {
-                       prev_sreg2 = -1;
+               mono_inst_set_src_registers (ins, sregs);
+
+               /* Sanity check */
+               /* Do this only for CAS for now */
+               for (j = 1; j < num_sregs; ++j) {
+                       int sreg = sregs [j];
+                       int dest_sreg = dest_sregs [j];
+
+                       if (j == 2 && dest_sreg != -1) {
+                               int k;
+
+                               g_assert (sreg == dest_sreg);
+
+                               for (k = 0; k < num_sregs; ++k) {
+                                       if (k != j)
+                                               g_assert (sregs [k] != dest_sreg);
+                               }
+                       }
                }
 
                /*if (reg_is_freeable (ins->sreg1) && prev_sreg1 >= 0 && reginfo [prev_sreg1].born_in >= i) {
@@ -1868,6 +2183,8 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                MonoInst *fxch;
                int tmp;
 
+               g_assert (num_sregs <= 2);
+
                for (ins = bb->code; ins; ins = ins->next) {
                        spec = ins_get_spec (ins->opcode);
 
@@ -1900,8 +2217,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                        MONO_INST_NEW (cfg, fxch, OP_X86_FXCH);
                                        fxch->inst_imm = sp - 1 - i;
 
-                                       prev->next = fxch;
-                                       fxch->next = ins;
+                                       mono_bblock_insert_after_ins (bb, prev, fxch);
                                        prev = fxch;
 
                                        tmp = fpstack [sp - 1];
@@ -1915,8 +2231,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                MONO_INST_NEW (cfg, fxch, OP_X86_FXCH);
                                fxch->inst_imm = 1;
 
-                               prev->next = fxch;
-                               fxch->next = ins;
+                               mono_bblock_insert_after_ins (bb, prev, fxch);
                                prev = fxch;
 
                                tmp = fpstack [sp - 1];
@@ -1940,8 +2255,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                        MONO_INST_NEW (cfg, fxch, OP_X86_FXCH);
                                        fxch->inst_imm = sp - 1 - i;
 
-                                       prev->next = fxch;
-                                       fxch->next = ins;
+                                       mono_bblock_insert_after_ins (bb, prev, fxch);
                                        prev = fxch;
 
                                        tmp = fpstack [sp - 1];
@@ -1968,8 +2282,7 @@ mono_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
                                        MONO_INST_NEW (cfg, fxch, OP_X86_FXCH);
                                        fxch->inst_imm = sp - 1 - i;
 
-                                       prev->next = fxch;
-                                       fxch->next = ins;
+                                       mono_bblock_insert_after_ins (bb, prev, fxch);
                                        prev = fxch;
 
                                        tmp = fpstack [sp - 1];
@@ -2015,7 +2328,6 @@ CompRelation
 mono_opcode_to_cond (int opcode)
 {
        switch (opcode) {
-       case CEE_BEQ:
        case OP_CEQ:
        case OP_IBEQ:
        case OP_ICEQ:
@@ -2028,7 +2340,6 @@ mono_opcode_to_cond (int opcode)
        case OP_CMOV_IEQ:
        case OP_CMOV_LEQ:
                return CMP_EQ;
-       case CEE_BNE_UN:
        case OP_IBNE_UN:
        case OP_LBNE_UN:
        case OP_FBNE_UN:
@@ -2037,21 +2348,18 @@ mono_opcode_to_cond (int opcode)
        case OP_CMOV_INE_UN:
        case OP_CMOV_LNE_UN:
                return CMP_NE;
-       case CEE_BLE:
        case OP_IBLE:
        case OP_LBLE:
        case OP_FBLE:
        case OP_CMOV_ILE:
        case OP_CMOV_LLE:
                return CMP_LE;
-       case CEE_BGE:
        case OP_IBGE:
        case OP_LBGE:
        case OP_FBGE:
        case OP_CMOV_IGE:
        case OP_CMOV_LGE:
                return CMP_GE;
-       case CEE_BLT:
        case OP_CLT:
        case OP_IBLT:
        case OP_ICLT:
@@ -2064,7 +2372,6 @@ mono_opcode_to_cond (int opcode)
        case OP_CMOV_ILT:
        case OP_CMOV_LLT:
                return CMP_LT;
-       case CEE_BGT:
        case OP_CGT:
        case OP_IBGT:
        case OP_ICGT:
@@ -2078,7 +2385,6 @@ mono_opcode_to_cond (int opcode)
        case OP_CMOV_LGT:
                return CMP_GT;
 
-       case CEE_BLE_UN:
        case OP_IBLE_UN:
        case OP_LBLE_UN:
        case OP_FBLE_UN:
@@ -2087,14 +2393,12 @@ mono_opcode_to_cond (int opcode)
        case OP_CMOV_ILE_UN:
        case OP_CMOV_LLE_UN:
                return CMP_LE_UN;
-       case CEE_BGE_UN:
        case OP_IBGE_UN:
        case OP_LBGE_UN:
        case OP_FBGE_UN:
        case OP_CMOV_IGE_UN:
        case OP_CMOV_LGE_UN:
                return CMP_GE_UN;
-       case CEE_BLT_UN:
        case OP_CLT_UN:
        case OP_IBLT_UN:
        case OP_ICLT_UN:
@@ -2107,7 +2411,6 @@ mono_opcode_to_cond (int opcode)
        case OP_CMOV_ILT_UN:
        case OP_CMOV_LLT_UN:
                return CMP_LT_UN;
-       case CEE_BGT_UN:
        case OP_CGT_UN:
        case OP_IBGT_UN:
        case OP_ICGT_UN:
@@ -2159,9 +2462,7 @@ mono_negate_cond (CompRelation cond)
 CompType
 mono_opcode_to_type (int opcode, int cmp_opcode)
 {
-       if ((opcode >= CEE_BEQ) && (opcode <= CEE_BLT_UN))
-               return CMP_TYPE_L;
-       else if ((opcode >= OP_CEQ) && (opcode <= OP_CLT_UN))
+       if ((opcode >= OP_CEQ) && (opcode <= OP_CLT_UN))
                return CMP_TYPE_L;
        else if ((opcode >= OP_IBEQ) && (opcode <= OP_IBLT_UN))
                return CMP_TYPE_I;
@@ -2181,7 +2482,6 @@ mono_opcode_to_type (int opcode, int cmp_opcode)
                switch (cmp_opcode) {
                case OP_ICOMPARE:
                case OP_ICOMPARE_IMM:
-               case OP_LCOMPARE_IMM:
                        return CMP_TYPE_I;
                default:
                        return CMP_TYPE_L;
@@ -2192,6 +2492,8 @@ mono_opcode_to_type (int opcode, int cmp_opcode)
        }
 }
 
+#endif /* DISABLE_JIT */
+
 gboolean
 mono_is_regsize_var (MonoType *t)
 {
@@ -2211,7 +2513,7 @@ mono_is_regsize_var (MonoType *t)
        case MONO_TYPE_U:
        case MONO_TYPE_PTR:
        case MONO_TYPE_FNPTR:
-#if SIZEOF_VOID_P == 8
+#if SIZEOF_REGISTER == 8
        case MONO_TYPE_I8:
        case MONO_TYPE_U8:
 #endif
@@ -2228,10 +2530,13 @@ mono_is_regsize_var (MonoType *t)
                return FALSE;
        case MONO_TYPE_VALUETYPE:
                return FALSE;
+       default:
+               return FALSE;
        }
-       return FALSE;
 }
 
+#ifndef DISABLE_JIT
+
 /*
  * mono_peephole_ins:
  *
@@ -2263,6 +2568,8 @@ mono_peephole_ins (MonoBasicBlock *bb, MonoInst *ins)
                 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
                 * OP_MOVE reg1, reg2
                 */
+               if (last_ins && last_ins->opcode == OP_GC_LIVENESS_DEF)
+                       last_ins = last_ins->prev;
                if (last_ins &&
                        (((ins->opcode == OP_LOADI4_MEMBASE) && (last_ins->opcode == OP_STOREI4_MEMBASE_REG)) ||
                         ((ins->opcode == OP_LOAD_MEMBASE) && (last_ins->opcode == OP_STORE_MEMBASE_REG))) &&
@@ -2351,7 +2658,7 @@ mono_peephole_ins (MonoBasicBlock *bb, MonoInst *ins)
                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 SIZEOF_VOID_P == 8
+#if SIZEOF_REGISTER == 8
                        ins->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? OP_PCONV_TO_I2 : OP_PCONV_TO_U2;
 #else
                        /* The definition of OP_PCONV_TO_U2 is wrong */
@@ -2377,7 +2684,7 @@ mono_peephole_ins (MonoBasicBlock *bb, MonoInst *ins)
                 * OP_MOVE sreg, dreg 
                 * OP_MOVE dreg, sreg
                 */
-               if (last_ins && last_ins->opcode == OP_MOVE &&
+               if (last_ins && last_ins->opcode == ins->opcode &&
                        ins->sreg1 == last_ins->dreg &&
                        ins->dreg == last_ins->sreg1) {
                        MONO_DELETE_INS (bb, ins);
@@ -2389,3 +2696,4 @@ mono_peephole_ins (MonoBasicBlock *bb, MonoInst *ins)
        }
 }
 
+#endif /* DISABLE_JIT */