--- /sys/src/9/port/fpi.h	Thu Jun 13 22:35:49 2013
+++ /sys/src/9/port/fpi.h	Thu Jun 20 23:47:44 2013
@@ -1,10 +1,15 @@
+#ifndef nil
+#include <u.h>
+#endif
+
 typedef long Word;
 typedef long long Vlong;
 typedef unsigned long Single;
-typedef struct {
-	unsigned long l;
-	unsigned long h;
-} Double;
+
+/* use u.h's FPdbleword */
+#define Double	FPdbleword
+#define h hi
+#define l lo
 
 enum {
 	FractBits	= 28,
--- /sys/src/9/teg2/syscall.c	Tue Apr  9 23:55:59 2013
+++ /sys/src/9/teg2/syscall.c	Fri Jun 21 21:51:13 2013
@@ -80,11 +80,11 @@
 		qunlock(&up->debug);
 
 		splhi();
-		nf->arg1 = nf->msg;
-		nf->arg0 = &nf->ureg;
+		nf->arg1 = nf->msg;		/* arg 1 is string */
+		nf->arg0 = &nf->ureg;		/* arg 0 XX(FP) is ureg* */
 		nf->ip = 0;
 		cur->sp = PTR2UINT(nf);
-		cur->r0 = PTR2UINT(nf->arg0);
+		cur->r0 = PTR2UINT(nf->arg0);	/* arg 0 in reg is ureg* */
 		break;
 	default:
 		pprint("unknown noted arg %#p\n", arg0);
@@ -168,9 +168,9 @@
 	nf->old = up->ureg;
 	up->ureg = nf;
 	memmove(nf->msg, up->note[0].msg, ERRMAX);
-	nf->arg1 = nf->msg;
-	nf->arg0 = &nf->ureg;
-	ureg->r0 = PTR2UINT(nf->arg0);
+	nf->arg1 = nf->msg;			/* arg 1 is string */
+	nf->arg0 = &nf->ureg;			/* arg 0 XX(FP) is ureg* */
+	ureg->r0 = PTR2UINT(nf->arg0);		/* arg 0 in r0 is ureg* */
 	nf->ip = 0;
 
 	ureg->sp = sp;
--- /sys/src/cmd/vl/l.h	Wed Mar 27 00:18:09 2013
+++ /sys/src/cmd/vl/l.h	Sun Jun 23 06:46:06 2013
@@ -251,6 +251,7 @@
 	Count	mfrom;
 	Count	page;
 	Count	jump;
+	Count	store;
 } nop;
 
 extern	char*	anames[];
--- /sys/src/cmd/vl/noop.c	Wed Jun 19 04:21:50 2013
+++ /sys/src/cmd/vl/noop.c	Sun Jun 23 06:46:35 2013
@@ -1,5 +1,184 @@
 #include	"l.h"
 
+/*
+ * flag: insert nops to prevent three consecutive stores.
+ * workaround for 24k erratum #48, costs about 10% in text space,
+ * so only enable this if you need it.  a test case is "hoc -e '7^6'".
+ */
+enum {
+	Mips24k	= 0,
+};
+
+static int
+isdblwrdmov(Prog *p)
+{
+	if(p == nil)
+		return 0;
+	switch(p->as){
+	case AMOVD:
+	case AMOVDF:
+	case AMOVDW:
+	case AMOVFD:
+	case AMOVWD:
+	case AMOVV:
+	case AMOVVL:
+	case AMOVVR:
+	case AMOVFV:
+	case AMOVDV:
+	case AMOVVF:
+	case AMOVVD:
+		return 1;
+	}
+	return 0;
+}
+
+static int
+ismove(Prog *p)
+{
+	if(p == nil)
+		return 0;
+	switch(p->as){
+	case AMOVB:
+	case AMOVBU:
+	case AMOVF:
+	case AMOVFW:
+	case AMOVH:
+	case AMOVHU:
+	case AMOVW:
+	case AMOVWF:
+	case AMOVWL:
+	case AMOVWR:
+	case AMOVWU:
+		return 1;
+	}
+	if(isdblwrdmov(p))
+		return 1;
+	return 0;
+}
+
+static int
+isstore(Prog *p)
+{
+	if(p == nil)
+		return 0;
+	if(ismove(p))
+		switch(p->to.type) {
+		case D_OREG:
+		case D_EXTERN:
+		case D_STATIC:
+		case D_AUTO:
+		case D_PARAM:
+			return 1;
+		}
+	return 0;
+}
+
+static int
+iscondbranch(Prog *p)
+{
+	if(p == nil)
+		return 0;
+	switch(p->as){
+	case ABEQ:
+	case ABFPF:
+	case ABFPT:
+	case ABGEZ:
+	case ABGEZAL:
+	case ABGTZ:
+	case ABLEZ:
+	case ABLTZ:
+	case ABLTZAL:
+	case ABNE:
+		return 1;
+	}
+	return 0;
+}
+
+static int
+isbranch(Prog *p)
+{
+	if(p == nil)
+		return 0;
+	switch(p->as){
+	case AJAL:
+	case AJMP:
+	case ARET:
+	case ARFE:
+		return 1;
+	}
+	if(iscondbranch(p))
+		return 1;
+	return 0;
+}
+
+/*
+ * workaround for 24k erratum #48, costs about 0.5% in space.
+ * inserts a NOP before the last of 3 consecutive stores.
+ * double-word stores complicate things.
+ */
+static int
+no3stores(Prog *p)
+{
+	Prog *p1;
+
+	if(!isstore(p))
+		return 0;
+	p1 = p->link;
+	if(!isstore(p1))
+		return 0;
+	if(isdblwrdmov(p) || isdblwrdmov(p1)) {
+		p->mark |= LABEL|SYNC;
+		addnop(p);
+		nop.store.count++;
+		nop.store.outof++;
+		return 1;
+	}
+	if(isstore(p1->link)) {
+		p1->mark |= LABEL|SYNC;
+		addnop(p1);
+		nop.store.count++;
+		nop.store.outof++;
+		return 1;
+	}
+	return 0;
+}
+
+/*
+ * keep stores out of branch delay slots.
+ * this is costly in space (the other 9.5%), but makes no3stores effective.
+ * there is undoubtedly a better way to do this.
+ */
+void
+storesnosched(void)
+{
+	Prog *p;
+
+	for(p = firstp; p != P; p = p->link)
+		if(isstore(p))
+			p->mark |= NOSCHED; /* keep stores out of delay slots */
+}
+
+void
+triplestorenops(void)
+{
+	Prog *p, *p1;
+
+	for(p = firstp; p != P; p = p1) {
+		p1 = p->link;
+//		if (p->mark & NOSCHED)
+//			continue;
+		if(ismove(p))
+			no3stores(p);
+		else if (isbranch(p))
+			/*
+			 * can't ignore delay slot of a conditional branch;
+			 * the branch could fail and fall through.
+			 */
+			if (!iscondbranch(p) && p1)
+				p1 = p1->link;	/* skip its delay slot */
+	}
+}
+
 void
 noops(void)
 {
@@ -348,6 +527,8 @@
 			break;
 		}
 	}
+	if (Mips24k)
+		storesnosched();
 
 	curtext = P;
 	q = P;		/* p - 1 */
@@ -388,6 +569,9 @@
 		}
 		q = p;
 	}
+
+	if (Mips24k)
+		triplestorenops();
 }
 
 void