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.

4 files changed, 40 insertions, 19 deletions

diff --git a/sys/src/cmd/8c/cgen.c b/sys/src/cmd/8c/cgen.c
index 35559bc15..1900f3cbc 100644
--- a/sys/src/cmd/8c/cgen.c
+++ b/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:
diff --git a/sys/src/cmd/8c/peep.c b/sys/src/cmd/8c/peep.c
index 8f0fae88a..482f17a65 100644
--- a/sys/src/cmd/8c/peep.c
+++ b/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:
-
-		case AREP:
-		case AREPN:
-
-		case ACWD:
-		case ACDQ:
-
-		case ASTOSB:
-		case ASTOSL:
-		case AMOVSB:
-		case AMOVSL:
-		case AFSTSW:
-			return 0;
+			if(p->from.type == D_CX && v1->type == D_CX)
+				return 0;
+			break;
 
 		case AORL:
 		case AANDL:
diff --git a/sys/src/cmd/8c/sgen.c b/sys/src/cmd/8c/sgen.c
index 9ba278c8c..6fc0ffba7 100644
--- a/sys/src/cmd/8c/sgen.c
+++ b/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:
diff --git a/sys/src/cmd/8c/txt.c b/sys/src/cmd/8c/txt.c
index 2974af4cc..9d64aa7e3 100644
--- a/sys/src/cmd/8c/txt.c
+++ b/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;