strconv: make FormatFloat slowpath a little faster

The relevant benchmark (on an Intel i7-4510U machine):

name                      old time/op  new time/op  delta
FormatFloat/Slowpath64-4  68.6µs ± 0%  44.1µs ± 2%  -35.71%  (p=0.000 n=13+15)

Change-Id: I67eb0e81ce74ed57752d0280059f91419f09e93b
Reviewed-on: https://go-review.googlesource.com/30099Reviewed-by: Robert Griesemer <gri@golang.org>

src/strconv/decimal.go

@@ -131,11 +131,13 @@ func rightShift(a *decimal, k uint) {
 	}
 	a.dp -= r - 1
 
+	var mask uint = (1 << k) - 1
+
 	// Pick up a digit, put down a digit.
 	for ; r < a.nd; r++ {
 		c := uint(a.d[r])
 		dig := n >> k
-		n -= dig << k
+		n &= mask
 		a.d[w] = byte(dig + '0')
 		w++
 		n = n*10 + c - '0'
@@ -144,7 +146,7 @@ func rightShift(a *decimal, k uint) {
 	// Put down extra digits.
 	for n > 0 {
 		dig := n >> k
-		n -= dig << k
+		n &= mask
 		if w < len(a.d) {
 			a.d[w] = byte(dig + '0')
 			w++

src/strconv/ftoa_test.go

@@ -208,6 +208,9 @@ var ftoaBenches = []struct {
 	{"64Fixed2", 123.456, 'e', 3, 64},
 	{"64Fixed3", 1.23456e+78, 'e', 3, 64},
 	{"64Fixed4", 1.23456e-78, 'e', 3, 64},
+
+	// Trigger slow path (see issue #15672).
+	{"Slowpath64", 622666234635.3213e-320, 'e', -1, 64},
 }
 
 func BenchmarkFormatFloat(b *testing.B) {