--- a/mc-code-spu.c	Mon Dec 06 16:23:42 2010 +0900
+++ b/mc-code-spu.c	Sat Sep 24 01:57:04 2011 +0900
@@ -127,20 +127,20 @@
 
 
 #define REG_ip   13
-#define REG_fp   0
+#define REG_fp   127
 #define REG_sp   1
 #define REG_VAR_BASE 3  
 #define REG_VAR_MIN  3
-#define REG_VAR_USER_MAX  127     /* at leat 6 tmp var */
-#define REG_VAR_MAX 127
+#define REG_VAR_USER_MAX  (127-8)     /* at leat 6 tmp var */
+#define REG_VAR_MAX (127-8)
 #define MIN_TMP_REG 3
-#define MAX_TMP_REG 127
+#define MAX_TMP_REG 126
 
 #define FREG_VAR_BASE 3
 #define FREG_VAR_MIN  3
-#define FREG_VAR_MAX  127
+#define FREG_VAR_MAX  REG_VAR_MAX
 #define MIN_TMP_FREG 3
-#define MAX_TMP_FREG 127
+#define MAX_TMP_FREG 126
 
 int MAX_REGISTER=127;             /* ARMのレジスタを10個まで使う*/
 #define  REAL_MAX_REGISTER 128    /* ARMのレジスタが32ということ*/
@@ -1026,7 +1026,7 @@
     printf("\tand\t%s,%s,%s\n",register_name(reg),register_name(reg),register_name(reg+1));
 }
 
-void
+static void
 cext(int sz, int sign, int from, int to)
 {
     if (sz==1 && !sign) {
@@ -1054,6 +1054,39 @@
     }
 }
 
+static void
+shuffle_offset(int sz, int offset, int base, int to)
+{
+    int tmp = get_register();
+    char *bn = register_name(base);
+    char *tn = register_name(tmp);
+
+    if (sz==1) {
+	printf("	cbd     %s %d(%s)\n",tn,offset, bn);
+	printf("	shufb   %s,%s,%s,%s\n",tn,fn,tn);
+    } else if (sz==1 && sign) {
+// signed char
+	printf("	rotqbyi %s,%s,13\n",fn,fn);
+	printf("	xsbh    %s,%s\n",fn,fn);
+	printf("	xshw    %s,%s\n",tn,fn);
+    } else if (sz==2 && !sign) {
+// unsigned short
+	int t = get_register();
+	char *tmp = register_name(t);
+	printf("	rotqbyi %s,%s,14\n",fn,fn);
+	printf("	ori     %s,%s,0\n",tmp,fn);
+	printf("	ila     %s,65535\n",fn,fn);
+	printf("	and     %s,%s,%s\n",tn,tmp,fn);
+	free_register(t);
+    } else if (sz==2 && !sign) {
+// singed short
+	printf("	rotqbyi	%s,%s,14\n",fn,fn);
+	printf("	xshw	%s,%s\n",tn,fn);
+    }
+    free_register(tmp);
+}
+
+
 void
 code_crlvar(int e2,int reg,int sign,int sz) {
     use_int(reg);

--- a/mc-codegen.c	Mon Dec 06 16:23:42 2010 +0900
+++ b/mc-codegen.c	Sat Sep 24 01:57:04 2011 +0900
@@ -1042,7 +1042,7 @@
 /*     source (after) list2(tag,disp)                       */
 /* source list    list3(e,cdr,sz)                           */
 
-#define DEBUG_PARALLEL_ASSIGN 0
+#define DEBUG_PARALLEL_ASSIGN 1
 
 static int is_writable(int);
 
@@ -1193,7 +1193,7 @@
 	for(target0=*target;target0; target0=cadr(target0)) {
 	    t=car(target0); s=cadddr(target0);
 	    sz=size(ty=caddr(target0)); 
-	    if(car(t)==car(s) && cadr(t)==cadr(s)) {
+	    if(is_same_type(t,s) && cadr(t)==cadr(s)) {
 		/*書き込み先が自分自身*/
 #if DEBUG_PARALLEL_ASSIGN
 if (lsrc) printf("## remove same %d ty %d+%d sz %d\n",car(t),ty,cadr(t),sz);
@@ -1295,8 +1295,10 @@
 {
     int ce1=car(e1);
     int ce2=car(e2);
+    if (ce1==ce2) return 1;
     if (ce1==LVAR) {
 	switch(ce2) {
+	    case LRLVAR: case LURLVAR:  return lp64;
 	    case RLVAR: case CRLVAR: case FRLVAR: case DRLVAR:
 	    case SRLVAR: case SURLVAR: case CURLVAR: case LVAR:
 	    case URLVAR:
@@ -1304,8 +1306,9 @@
 	}
     } else if (ce2==LVAR) {
 	switch(ce1) {
+	    case LRLVAR: case LURLVAR:  return lp64;
 	    case RLVAR: case CRLVAR: case FRLVAR: case DRLVAR:
-	    case SRLVAR: case SURLVAR: case CURLVAR: case LRLVAR: case LVAR:
+	    case SRLVAR: case SURLVAR: case CURLVAR: case LVAR:
 	    case URLVAR:
 	    return 1;
 	}
@@ -1327,6 +1330,10 @@
 	    return 1;
 	}
 #endif
+    } else if (ce1==REGISTER) {
+	if (lp64 && ce2==LREGISTER) return 1;
+    } else if (ce2==REGISTER) {
+	if (lp64 && ce1==LREGISTER) return 1;
     }
     return 0;
 }
@@ -1337,7 +1344,7 @@
     switch(car(e1)) {
     case GVAR :
     case LVAR :
-    case RLVAR :        // this is wrong, but ia32 generates this.
+    case RLVAR :  case LRLVAR: case URLVAR: case LURLVAR:     // this is wrong, but ia32 generates this.
     case REGISTER :
 	return size_of_int;
     case FREGISTER :