plateau in divide by a constant

still to do - overflow, mod

R=rsc
OCL=32927
CL=32927

src/cmd/6g/ggen.c

@@ -657,7 +657,6 @@ cgen_div(int op, Node *nl, Node *nr, Node *res)
 	return;
 
 divbymul:
-goto longdiv;
 	switch(simtype[nl->type->etype]) {
 	default:
 		goto longdiv;
@@ -678,7 +677,29 @@ goto longdiv;
 			// todo fixup
 			break;
 		}
-		break;
+
+		savex(D_AX, &ax, &oldax, res, nl->type);
+		savex(D_DX, &dx, &olddx, res, nl->type);
+		savex(D_CX, &cx, &oldcx, res, nl->type);
+
+		regalloc(&n1, nl->type, N);
+		cgen(nl, &n1);				// num -> reg(n1)
+
+		nodconst(&n2, nl->type, m.um);
+		gmove(&n2, &ax);			// const->ax
+
+		gins(optoas(OHMUL, nl->type), &n1, N);	// imul reg
+
+		nodconst(&n2, nl->type, m.s);
+		gins(optoas(ORSH, nl->type), &n2, &dx);	// shift dx
+
+		regfree(&n1);
+		gmove(&dx, res);
+
+		restx(&ax, &oldax);
+		restx(&dx, &olddx);
+		restx(&cx, &oldcx);
+		return;
 
 	case TINT16:
 	case TINT32:
@@ -707,7 +728,7 @@ goto longdiv;
 		nodconst(&n2, nl->type, m.sm);
 		gmove(&n2, &ax);			// const->ax
 
-		gins(optoas(OMUL, nl->type), &n1, N);	// imul reg
+		gins(optoas(OHMUL, nl->type), &n1, N);	// imul reg
 
 		nodconst(&n2, nl->type, m.s);
 		gins(optoas(ORSH, nl->type), &n2, &dx);	// shift dx

src/cmd/6g/gsubr.c

@@ -1473,28 +1473,50 @@ optoas(int op, Type *t)
 		a = ASARQ;
 		break;
 
+	case CASE(OHMUL, TINT8):
 	case CASE(OMUL, TINT8):
 	case CASE(OMUL, TUINT8):
 		a = AIMULB;
 		break;
 
+	case CASE(OHMUL, TINT16):
 	case CASE(OMUL, TINT16):
 	case CASE(OMUL, TUINT16):
 		a = AIMULW;
 		break;
 
+	case CASE(OHMUL, TINT32):
 	case CASE(OMUL, TINT32):
 	case CASE(OMUL, TUINT32):
 	case CASE(OMUL, TPTR32):
 		a = AIMULL;
 		break;
 
+	case CASE(OHMUL, TINT64):
 	case CASE(OMUL, TINT64):
 	case CASE(OMUL, TUINT64):
 	case CASE(OMUL, TPTR64):
 		a = AIMULQ;
 		break;
 
+	case CASE(OHMUL, TUINT8):
+		a = AMULB;
+		break;
+
+	case CASE(OHMUL, TUINT16):
+		a = AMULW;
+		break;
+
+	case CASE(OHMUL, TUINT32):
+	case CASE(OHMUL, TPTR32):
+		a = AMULL;
+		break;
+
+	case CASE(OHMUL, TUINT64):
+	case CASE(OHMUL, TPTR64):
+		a = AMULQ;
+		break;
+
 	case CASE(OMUL, TFLOAT32):
 		a = AMULSS;
 		break;
@@ -1930,8 +1952,8 @@ void
 smagic(Magic *m)
 {
 	int p;
-	uint64 ad, anc, delta, q1, r1, q2, r2, t, two31;
-	uint64 mask;
+	uint64 ad, anc, delta, q1, r1, q2, r2, t;
+	uint64 mask, two31;
 
 	m->bad = 0;
 	switch(m->w) {
@@ -2013,6 +2035,8 @@ smagic(Magic *m)
 	}
 
 	m->sm = q2+1;
+	if(m->sm & two31)
+		m->sm |= ~mask;
 	m->s = p-m->w;
 }
 
@@ -2020,8 +2044,8 @@ void
 umagic(Magic *m)
 {
 	int p;
-	uint64 nc, delta, q1, r1, q2, r2, two31;
-	uint64 mask;
+	uint64 nc, delta, q1, r1, q2, r2;
+	uint64 mask, two31;
 
 	m->bad = 0;
 	m->ua = 0;

src/cmd/gc/go.h

@@ -343,7 +343,7 @@ enum
 	OKEY, OPARAM,
 	OLEN,
 	OMAKE, OMAKECHAN, OMAKEMAP, OMAKESLICE,
-	OMUL, ODIV, OMOD, OLSH, ORSH, OAND, OANDNOT,
+	OMUL, ODIV, OMOD, OLSH, ORSH, OHMUL, OAND, OANDNOT,
 	ONEW,
 	ONOT, OCOM, OPLUS, OMINUS,
 	OOROR,