runtime: implement 64 bit division in Go

LGTM=rsc, dave
R=minux, rsc, remyoudompheng, dave
CC=golang-codereviews
https://golang.org/cl/133790043

src/pkg/runtime/vlop_386.s

@@ -29,9 +29,9 @@
  * C runtime for 64-bit divide.
  */
 
-// _mul64x32(r *uint64, a uint64, b uint32) uint32
+// runtime·_mul64x32(r *uint64, a uint64, b uint32) uint32
 // sets *r = low 64 bits of 96-bit product a*b; returns high 32 bits.
-TEXT _mul64by32(SB), NOSPLIT, $0
+TEXT runtime·_mul64by32(SB), NOSPLIT, $0
 	MOVL	r+0(FP), CX
 	MOVL	a+4(FP), AX
 	MULL	b+12(FP)
@@ -46,7 +46,7 @@ TEXT _mul64by32(SB), NOSPLIT, $0
 	MOVL	AX, ret+16(FP)
 	RET
 
-TEXT _div64by32(SB), NOSPLIT, $0
+TEXT runtime·_div64by32(SB), NOSPLIT, $0
 	MOVL	r+12(FP), CX
 	MOVL	a+0(FP), AX
 	MOVL	a+4(FP), DX

src/pkg/runtime/vlop_arm.s

@@ -294,3 +294,12 @@ out:
 	MOVW	12(R13), R(s)
 	MOVW	16(R13), R(M)
 	RET
+
+// _mul64by32 and _div64by32 not implemented on arm
+TEXT runtime·_mul64by32(SB), NOSPLIT, $0
+	MOVW	$0, R0
+	MOVW	(R0), R1 // crash
+
+TEXT runtime·_div64by32(SB), NOSPLIT, $0
+	MOVW	$0, R0
+	MOVW	(R0), R1 // crash

src/pkg/runtime/vlrt.c

@@ -34,7 +34,6 @@
  * to generate the code directly now.  Find and remove.
  */
 
-void	runtime·panicstring(char*);
 void	runtime·panicdivide(void);
 
 typedef	unsigned long	ulong;
@@ -182,8 +181,8 @@ _v2f(Vlong x)
 	return _v2d(x);
 }
 
-ulong	_div64by32(Vlong, ulong, ulong*);
-int	_mul64by32(Vlong*, Vlong, ulong);
+ulong	runtime·_div64by32(Vlong, ulong, ulong*);
+int	runtime·_mul64by32(Vlong*, Vlong, ulong);
 
 static void
 slowdodiv(Vlong num, Vlong den, Vlong *q, Vlong *r)
@@ -277,7 +276,7 @@ dodiv(Vlong num, Vlong den, Vlong *qp, Vlong *rp)
 	if(den.hi != 0){
 		q.hi = 0;
 		n = num.hi/den.hi;
-		if(_mul64by32(&x, den, n) || x.hi > num.hi || (x.hi == num.hi && x.lo > num.lo))
+		if(runtime·_mul64by32(&x, den, n) || x.hi > num.hi || (x.hi == num.hi && x.lo > num.lo))
 			slowdodiv(num, den, &q, &r);
 		else {
 			q.lo = n;
@@ -292,7 +291,7 @@ dodiv(Vlong num, Vlong den, Vlong *qp, Vlong *rp)
 		} else {
 			q.hi = 0;
 		}
-		q.lo = _div64by32(num, den.lo, &r.lo);
+		q.lo = runtime·_div64by32(num, den.lo, &r.lo);
 		r.hi = 0;
 	}
 	if(qp) {
@@ -322,12 +321,6 @@ _divvu(Vlong n, Vlong d)
 	return q;
 }
 
-Vlong
-runtime·uint64div(Vlong n, Vlong d)
-{
-	return _divvu(n, d);
-}
-
 Vlong
 _modvu(Vlong n, Vlong d)
 {
@@ -344,12 +337,6 @@ _modvu(Vlong n, Vlong d)
 	return r;
 }
 
-Vlong
-runtime·uint64mod(Vlong n, Vlong d)
-{
-	return _modvu(n, d);
-}
-
 static void
 vneg(Vlong *v)
 {
@@ -394,12 +381,6 @@ _divv(Vlong n, Vlong d)
 	return q;
 }
 
-Vlong
-runtime·int64div(Vlong n, Vlong d)
-{
-	return _divv(n, d);
-}
-
 Vlong
 _modv(Vlong n, Vlong d)
 {
@@ -432,12 +413,6 @@ _modv(Vlong n, Vlong d)
 	return r;
 }
 
-Vlong
-runtime·int64mod(Vlong n, Vlong d)
-{
-	return _modv(n, d);
-}
-
 Vlong
 _rshav(Vlong a, int b)
 {

src/pkg/runtime/vlrt.go

@@ -23,13 +23,14 @@
 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 // THE SOFTWARE.
 
-// +build arm
+// +build arm 386
 
 package runtime
 
 import "unsafe"
 
 const (
+	sign32 = 1 << (32 - 1)
 	sign64 = 1 << (64 - 1)
 )
 
@@ -105,3 +106,155 @@ func _d2v(y *uint64, d float64) {
 
 	*y = uint64(yhi)<<32 | uint64(ylo)
 }
+
+func uint64div(n, d uint64) uint64 {
+	// Check for 32 bit operands
+	if uint32(n>>32) == 0 && uint32(d>>32) == 0 {
+		if uint32(d) == 0 {
+			panicdivide()
+		}
+		return uint64(uint32(n) / uint32(d))
+	}
+	q, _ := dodiv(n, d)
+	return q
+}
+
+func uint64mod(n, d uint64) uint64 {
+	// Check for 32 bit operands
+	if uint32(n>>32) == 0 && uint32(d>>32) == 0 {
+		if uint32(d) == 0 {
+			panicdivide()
+		}
+		return uint64(uint32(n) % uint32(d))
+	}
+	_, r := dodiv(n, d)
+	return r
+}
+
+func int64div(n, d int64) int64 {
+	// Check for 32 bit operands
+	if int64(int32(n)) == n && int64(int32(d)) == d {
+		if int32(n) == -0x80000000 && int32(d) == -1 {
+			// special case: 32-bit -0x80000000 / -1 = -0x80000000,
+			// but 64-bit -0x80000000 / -1 = 0x80000000.
+			return 0x80000000
+		}
+		if int32(d) == 0 {
+			panicdivide()
+		}
+		return int64(int32(n) / int32(d))
+	}
+
+	nneg := n < 0
+	dneg := d < 0
+	if nneg {
+		n = -n
+	}
+	if dneg {
+		d = -d
+	}
+	uq, _ := dodiv(uint64(n), uint64(d))
+	q := int64(uq)
+	if nneg != dneg {
+		q = -q
+	}
+	return q
+}
+
+func int64mod(n, d int64) int64 {
+	// Check for 32 bit operands
+	if int64(int32(n)) == n && int64(int32(d)) == d {
+		if int32(d) == 0 {
+			panicdivide()
+		}
+		return int64(int32(n) % int32(d))
+	}
+
+	nneg := n < 0
+	if nneg {
+		n = -n
+	}
+	if d < 0 {
+		d = -d
+	}
+	_, ur := dodiv(uint64(n), uint64(d))
+	r := int64(ur)
+	if nneg {
+		r = -r
+	}
+	return r
+}
+
+//go:noescape
+func _mul64by32(lo64 *uint64, a uint64, b uint32) (hi32 uint32)
+
+//go:noescape
+func _div64by32(a uint64, b uint32, r *uint32) (q uint32)
+
+func dodiv(n, d uint64) (q, r uint64) {
+	if GOARCH == "arm" {
+		// arm doesn't have a division instruction, so
+		// slowdodiv is the best that we can do.
+		// TODO: revisit for arm64.
+		return slowdodiv(n, d)
+	}
+
+	if d > n {
+		return 0, n
+	}
+
+	if uint32(d>>32) != 0 {
+		t := uint32(n>>32) / uint32(d>>32)
+		var lo64 uint64
+		hi32 := _mul64by32(&lo64, d, t)
+		if hi32 != 0 || lo64 > n {
+			return slowdodiv(n, d)
+		}
+		return uint64(t), n - lo64
+	}
+
+	// d is 32 bit
+	var qhi uint32
+	if uint32(n>>32) >= uint32(d) {
+		if uint32(d) == 0 {
+			panicdivide()
+		}
+		qhi = uint32(n>>32) / uint32(d)
+		n -= uint64(uint32(d)*qhi) << 32
+	} else {
+		qhi = 0
+	}
+
+	var rlo uint32
+	qlo := _div64by32(n, uint32(d), &rlo)
+	return uint64(qhi)<<32 + uint64(qlo), uint64(rlo)
+}
+
+func slowdodiv(n, d uint64) (q, r uint64) {
+	if d == 0 {
+		panicdivide()
+	}
+
+	// Set up the divisor and find the number of iterations needed.
+	capn := n
+	if n >= sign64 {
+		capn = sign64
+	}
+	i := 0
+	for d < capn {
+		d <<= 1
+		i++
+	}
+
+	for ; i >= 0; i-- {
+		q <<= 1
+		if n >= d {
+			n -= d
+			q |= 1
+		}
+		d >>= 1
+	}
+	return q, n
+}
+
+func panicdivide()