8c, 6c: native ROL (cyclic shift) instruction support, improve peephole optimizers

introduce rolor() function to subsitute (a << c) | (a >> (bits(a) - c))
with (a <<< c) where <<< is cyclic rotation and c is constant.
this almost doubles the speed of chacha encryption of 386 and amd64.

the peephole optimizer used to stop when it hit a shift or rol
instruction when attempting to eleminate moves by register
substitution. but we do not have to as long as the shift count
operand is not CX (which cannot be substituted) and CX is not
a subject for substitution.

sys/src/cmd/6c/cgen.c

@@ -159,6 +159,7 @@ cgen(Node *n, Node *nn)
 		regfree(&nod);
 		break;
 
+	case OROL:
 	case OLSHR:
 	case OASHL:
 	case OASHR:

sys/src/cmd/6c/peep.c

@@ -370,15 +370,11 @@ subprop(Reg *r0)
 			break;
 		p = r->prog;
 		switch(p->as) {
-		case ACALL:
-			return 0;
-
 		case AIMULL:
 		case AIMULQ:
 		case AIMULW:
 			if(p->to.type != D_NONE)
 				break;
-
 		case ADIVB:
 		case ADIVL:
 		case ADIVQ:
@@ -393,6 +389,19 @@ subprop(Reg *r0)
 		case AMULQ:
 		case AMULW:
 
+		case ACWD:
+		case ACDQ:
+		case ACQO:
+
+		case AREP:
+		case AREPN:
+		case ALOOP:
+		case ALOOPEQ:
+		case ALOOPNE:
+
+		case ACALL:
+			return 0;
+
 		case AROLB:
 		case AROLL:
 		case AROLQ:
@@ -417,14 +426,9 @@ subprop(Reg *r0)
 		case ASHRL:
 		case ASHRQ:
 		case ASHRW:
-
+			if(p->from.type == D_CX && v1->type == D_CX)
-		case AREP:
-		case AREPN:
-
-		case ACWD:
-		case ACDQ:
-		case ACQO:
 				return 0;
+			break;
 
 		case AORL:
 		case AORQ:

sys/src/cmd/6c/sgen.c

@@ -120,7 +120,6 @@ xcom(Node *n)
 				*l = *(n->left);
 				l->xoffset += r->vconst;
 				n->left = l;
-				r = n->right;
 				goto brk;
 			}
 			break;
@@ -212,7 +211,6 @@ xcom(Node *n)
 		if(g >= 0) {
 			n->left = r;
 			n->right = l;
-			l = r;
 			r = n->right;
 		}
 		g = vlog(r);
@@ -288,6 +286,12 @@ xcom(Node *n)
 		indexshift(n);
 		break;
 
+	case OOR:
+		xcom(l);
+		xcom(r);
+		rolor(n);
+		break;
+
 	default:
 		if(l != Z)
 			xcom(l);
@@ -298,6 +302,8 @@ xcom(Node *n)
 brk:
 	if(n->addable >= 10)
 		return;
+	l = n->left;
+	r = n->right;
 	if(l != Z)
 		n->complex = l->complex;
 	if(r != Z) {
@@ -344,6 +350,7 @@ brk:
 		}
 		break;
 
+	case OROL:
 	case OLSHR:
 	case OASHL:
 	case OASHR:

sys/src/cmd/6c/txt.c

@@ -1305,6 +1305,16 @@ gopcode(int o, Type *ty, Node *f, Node *t)
 			a = ASALQ;
 		break;
 
+	case OROL:
+		a = AROLL;
+		if(et == TCHAR || et == TUCHAR)
+			a = AROLB;
+		if(et == TSHORT || et == TUSHORT)
+			a = AROLW;
+		if(et == TVLONG || et == TUVLONG || et == TIND)
+			a = AROLQ;
+		break;
+
 	case OFUNC:
 		a = ACALL;
 		break;

sys/src/cmd/8c/cgen.c

@@ -178,6 +178,7 @@ cgen(Node *n, Node *nn)
 		regfree(&nod);
 		break;
 
+	case OROL:
 	case OLSHR:
 	case OASHL:
 	case OASHR:

sys/src/cmd/8c/peep.c

@@ -264,9 +264,6 @@ subprop(Reg *r0)
 			break;
 		p = r->prog;
 		switch(p->as) {
-		case ACALL:
-			return 0;
-
 		case AIMULL:
 		case AIMULW:
 			if(p->to.type != D_NONE)
@@ -283,6 +280,23 @@ subprop(Reg *r0)
 		case AMULL:
 		case AMULW:
 
+		case AREP:
+		case AREPN:
+		case ALOOP:
+		case ALOOPNE:
+
+		case ACWD:
+		case ACDQ:
+
+		case ASTOSB:
+		case ASTOSL:
+		case AMOVSB:
+		case AMOVSL:
+		case AFSTSW:
+
+		case ACALL:
+			return 0;
+
 		case AROLB:
 		case AROLL:
 		case AROLW:
@@ -301,19 +315,9 @@ subprop(Reg *r0)
 		case ASHRB:
 		case ASHRL:
 		case ASHRW:
-
+			if(p->from.type == D_CX && v1->type == D_CX)
-		case AREP:
-		case AREPN:
-
-		case ACWD:
-		case ACDQ:
-
-		case ASTOSB:
-		case ASTOSL:
-		case AMOVSB:
-		case AMOVSL:
-		case AFSTSW:
 				return 0;
+			break;
 
 		case AORL:
 		case AANDL:

sys/src/cmd/8c/sgen.c

@@ -127,7 +127,6 @@ xcom(Node *n)
 				*l = *(n->left);
 				l->xoffset += r->vconst;
 				n->left = l;
-				r = n->right;
 				goto brk;
 			}
 			break;
@@ -219,7 +218,6 @@ xcom(Node *n)
 		if(g >= 0) {
 			n->left = r;
 			n->right = l;
-			l = r;
 			r = n->right;
 		}
 		g = vlog(r);
@@ -230,7 +228,7 @@ xcom(Node *n)
 			indexshift(n);
 			break;
 		}
-commute(n);
+		commute(n);
 		break;
 
 	case OASLDIV:
@@ -295,6 +293,13 @@ commute(n);
 		indexshift(n);
 		break;
 
+	case OOR:
+		xcom(l);
+		xcom(r);
+		if(typechl[n->type->etype])
+			rolor(n);
+		break;
+
 	default:
 		if(l != Z)
 			xcom(l);
@@ -305,6 +310,8 @@ commute(n);
 brk:
 	if(n->addable >= 10)
 		return;
+	l = n->left;
+	r = n->right;
 	if(l != Z)
 		n->complex = l->complex;
 	if(r != Z) {
@@ -349,6 +356,7 @@ brk:
 		}
 		break;
 
+	case OROL:
 	case OLSHR:
 	case OASHL:
 	case OASHR:

sys/src/cmd/8c/txt.c

@@ -1238,6 +1238,14 @@ gopcode(int o, Type *ty, Node *f, Node *t)
 			a = ASALW;
 		break;
 
+	case OROL:
+		a = AROLL;
+		if(et == TCHAR || et == TUCHAR)
+			a = AROLB;
+		if(et == TSHORT || et == TUSHORT)
+			a = AROLW;
+		break;
+
 	case OFUNC:
 		a = ACALL;
 		break;

sys/src/cmd/cc/cc.h

@@ -273,6 +273,7 @@ enum
 	OPROTO,
 	OREGISTER,
 	ORETURN,
+	OROL,
 	OSET,
 	OSIGN,
 	OSIZE,
@@ -694,6 +695,7 @@ int	log2(uvlong);
 int	vlog(Node*);
 int	topbit(ulong);
 void	simplifyshift(Node*);
+void	rolor(Node*);
 long	typebitor(long, long);
 void	diag(Node*, char*, ...);
 void	warn(Node*, char*, ...);

sys/src/cmd/cc/sub.c

@@ -903,6 +903,46 @@ if(debug['<'])prtree(n, "rewrite2");
 	n->left->op = o;
 }
 
+/*
+ * replace shift/or with rotate left
+ */
+void
+rolor(Node *n)
+{
+	Node *l, *r;
+
+	if(!typeu[n->type->etype])
+		return;
+
+	l = n->left;
+	r = n->right;
+	switch(l->op){
+	case OASHL:
+		if(r->op == OLSHR)
+			break;
+		return;
+	case OLSHR:
+		if(r->op == OASHL){
+			r = l;
+			l = n->right;
+			break;
+		}
+	default:
+		return;
+	}
+	if(l->right->op != OCONST || r->right->op != OCONST)
+		return;
+	if(vconst(l->right) + vconst(r->right) != ewidth[n->type->etype]*8)
+		return;
+	if(l->left->type != n->type || r->left->type != n->type)
+		return;
+	if(l->left->op != ONAME || r->left->op != ONAME || l->left->sym != r->left->sym)
+		return;
+
+	*n = *l;
+	n->op = OROL;
+}
+
 int
 side(Node *n)
 {
@@ -1473,6 +1513,7 @@ Init	onamesinit[] =
 	OPROTO,		0,	"PROTO",
 	OREGISTER,	0,	"REGISTER",
 	ORETURN,	0,	"RETURN",
+	OROL,		0,	"ROL",
 	OSET,		0,	"SET",
 	OSIGN,		0,	"SIGN",
 	OSIZE,		0,	"SIZE",