6g, 8g: divide corner case

Fixes #1772.

R=ken2
CC=golang-dev
https://golang.org/cl/4798062

src/cmd/6g/ggen.c

@@ -436,9 +436,6 @@ samereg(Node *a, Node *b)
 
 /*
  * generate division.
- * caller must set:
- *	ax = allocated AX register
- *	dx = allocated DX register
  * generates one of:
  *	res = nl / nr
  *	res = nl % nr
@@ -447,17 +444,35 @@ samereg(Node *a, Node *b)
 void
 dodiv(int op, Node *nl, Node *nr, Node *res)
 {
-	int a;
-	Node n3, n4;
+	int a, check;
+	Node n3, n4, n5;
 	Type *t;
 	Node ax, dx, oldax, olddx;
+	Prog *p1, *p2, *p3;
 
+	// Have to be careful about handling
+	// most negative int divided by -1 correctly.
+	// The hardware will trap.
+	// Also the byte divide instruction needs AH,
+	// which we otherwise don't have to deal with.
+	// Easiest way to avoid for int8, int16: use int32.
+	// For int32 and int64, use explicit test.
+	// Could use int64 hw for int32.
 	t = nl->type;
-	if(t->width == 1) {
+	check = 0;
+	if(issigned[t->etype]) {
+		check = 1;
+		if(isconst(nl, CTINT) && mpgetfix(nl->val.u.xval) != -1LL<<(t->width*8-1))
+			check = 0;
+		else if(isconst(nr, CTINT) && mpgetfix(nr->val.u.xval) != -1)
+			check = 0;
+	}
+	if(t->width < 4) {
 		if(issigned[t->etype])
 			t = types[TINT32];
 		else
 			t = types[TUINT32];
+		check = 0;
 	}
 	a = optoas(op, t);
 
@@ -473,6 +488,31 @@ dodiv(int op, Node *nl, Node *nr, Node *res)
 		savex(D_AX, &ax, &oldax, res, t);
 		cgen(nl, &ax);
 	}
+	p3 = P;
+	if(check) {
+		nodconst(&n4, t, -1);
+		gins(optoas(OCMP, t), &n3, &n4);
+		p1 = gbranch(optoas(ONE, t), T);
+		nodconst(&n4, t, -1LL<<(t->width*8-1));
+		if(t->width == 8) {
+			n5 = n4;
+			regalloc(&n4, t, N);
+			gins(AMOVQ, &n5, &n4);
+		}
+		gins(optoas(OCMP, t), &ax, &n4);
+		p2 = gbranch(optoas(ONE, t), T);
+		if(op == ODIV)
+			gmove(&n4, res);
+		if(t->width == 8)
+			regfree(&n4);
+		if(op == OMOD) {
+			nodconst(&n4, t, 0);
+			gmove(&n4, res);
+		}
+		p3 = gbranch(AJMP, T);
+		patch(p1, pc);
+		patch(p2, pc);
+	}
 	savex(D_DX, &dx, &olddx, res, t);
 	if(!issigned[t->etype]) {
 		nodconst(&n4, t, 0);
@@ -481,13 +521,14 @@ dodiv(int op, Node *nl, Node *nr, Node *res)
 		gins(optoas(OEXTEND, t), N, N);
 	gins(a, &n3, N);
 	regfree(&n3);
-
 	if(op == ODIV)
 		gmove(&ax, res);
 	else
 		gmove(&dx, res);
-	restx(&ax, &oldax);
 	restx(&dx, &olddx);
+	if(check)
+		patch(p3, pc);
+	restx(&ax, &oldax);
 }
 
 /*

src/cmd/8g/ggen.c

@@ -480,12 +480,40 @@ samereg(Node *a, Node *b)
  * according to op.
  */
 void
-dodiv(int op, Type *t, Node *nl, Node *nr, Node *res, Node *ax, Node *dx)
+dodiv(int op, Node *nl, Node *nr, Node *res, Node *ax, Node *dx)
 {
-	Node n1, t1, t2, nz;
+	int check;
+	Node n1, t1, t2, n4, nz;
+	Type *t;
+	Prog *p1, *p2, *p3;
 
-	tempname(&t1, nl->type);
-	tempname(&t2, nr->type);
+	// Have to be careful about handling
+	// most negative int divided by -1 correctly.
+	// The hardware will trap.
+	// Also the byte divide instruction needs AH,
+	// which we otherwise don't have to deal with.
+	// Easiest way to avoid for int8, int16: use int32.
+	// For int32 and int64, use explicit test.
+	// Could use int64 hw for int32.
+	t = nl->type;
+	check = 0;
+	if(issigned[t->etype]) {
+		check = 1;
+		if(isconst(nl, CTINT) && mpgetfix(nl->val.u.xval) != -1LL<<(t->width*8-1))
+			check = 0;
+		else if(isconst(nr, CTINT) && mpgetfix(nr->val.u.xval) != -1)
+			check = 0;
+	}
+	if(t->width < 4) {
+		if(issigned[t->etype])
+			t = types[TINT32];
+		else
+			t = types[TUINT32];
+		check = 0;
+	}
+
+	tempname(&t1, t);
+	tempname(&t2, t);
 	cgen(nl, &t1);
 	cgen(nr, &t2);
 
@@ -495,6 +523,24 @@ dodiv(int op, Type *t, Node *nl, Node *nr, Node *res, Node *ax, Node *dx)
 		regalloc(&n1, t, N);
 	gmove(&t2, &n1);
 	gmove(&t1, ax);
+	p3 = P;
+	if(check) {
+		nodconst(&n4, t, -1);
+		gins(optoas(OCMP, t), &n1, &n4);
+		p1 = gbranch(optoas(ONE, t), T);
+		nodconst(&n4, t, -1LL<<(t->width*8-1));
+		gins(optoas(OCMP, t), ax, &n4);
+		p2 = gbranch(optoas(ONE, t), T);
+		if(op == ODIV)
+			gmove(&n4, res);
+		if(op == OMOD) {
+			nodconst(&n4, t, 0);
+			gmove(&n4, res);
+		}
+		p3 = gbranch(AJMP, T);
+		patch(p1, pc);
+		patch(p2, pc);
+	}
 	if(!issigned[t->etype]) {
 		nodconst(&nz, t, 0);
 		gmove(&nz, dx);
@@ -507,6 +553,8 @@ dodiv(int op, Type *t, Node *nl, Node *nr, Node *res, Node *ax, Node *dx)
 		gmove(ax, res);
 	else
 		gmove(dx, res);
+	if(check)
+		patch(p3, pc);
 }
 
 static void
@@ -553,13 +601,13 @@ cgen_div(int op, Node *nl, Node *nr, Node *res)
 	if(is64(nl->type))
 		fatal("cgen_div %T", nl->type);
 
-	t = nl->type;
-	if(t->width == 1)
-		t = types[t->etype+2];	// int8 -> int16, uint8 -> uint16
-
+	if(issigned[nl->type->etype])
+		t = types[TINT32];
+	else
+		t = types[TUINT32];
 	savex(D_AX, &ax, &oldax, res, t);
 	savex(D_DX, &dx, &olddx, res, t);
-	dodiv(op, t, nl, nr, res, &ax, &dx);
+	dodiv(op, nl, nr, res, &ax, &dx);
 	restx(&dx, &olddx);
 	restx(&ax, &oldax);
 }

test/divide.go

@@ -0,0 +1,54 @@
+// $G $D/$F.go && $L $F.$A && ./$A.out
+
+// Copyright 2011 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// divide corner cases
+
+package main
+
+import "fmt"
+
+func f8(x, y, q, r int8) {
+	if t := x / y; t != q {
+		fmt.Printf("%d/%d = %d, want %d\n", x, y, t, q)
+	}
+	if t := x % y; t != r {
+		fmt.Printf("%d%%%d = %d, want %d\n", x, y, t, r)
+	}
+}
+
+func f16(x, y, q, r int16) {
+	if t := x / y; t != q {
+		fmt.Printf("%d/%d = %d, want %d\n", x, y, t, q)
+	}
+	if t := x % y; t != r {
+		fmt.Printf("%d%%%d = %d, want %d\n", x, y, t, r)
+	}
+}
+
+func f32(x, y, q, r int32) {
+	if t := x / y; t != q {
+		fmt.Printf("%d/%d = %d, want %d\n", x, y, t, q)
+	}
+	if t := x % y; t != r {
+		fmt.Printf("%d%%%d = %d, want %d\n", x, y, t, r)
+	}
+}
+
+func f64(x, y, q, r int64) {
+	if t := x / y; t != q {
+		fmt.Printf("%d/%d = %d, want %d\n", x, y, t, q)
+	}
+	if t := x % y; t != r {
+		fmt.Printf("%d%%%d = %d, want %d\n", x, y, t, r)
+	}
+}
+
+func main() {
+	f8(-1<<7, -1, -1<<7, 0)
+	f16(-1<<15, -1, -1<<15, 0)
+	f32(-1<<31, -1, -1<<31, 0)
+	f64(-1<<63, -1, -1<<63, 0)
+}