strconv: make Ftoa faster

Make code amenable to escape analysis
so that the decimal values do not escape.

benchmark                               old ns/op    new ns/op    delta
strconv_test.BenchmarkAtof64Decimal           229          233   +1.75%
strconv_test.BenchmarkAtof64Float             261          263   +0.77%
strconv_test.BenchmarkAtof64FloatExp         7760         7757   -0.04%
strconv_test.BenchmarkAtof64Big              3086         3053   -1.07%
strconv_test.BenchmarkFtoa64Decimal          6866         2629  -61.71%
strconv_test.BenchmarkFtoa64Float            7211         3064  -57.51%
strconv_test.BenchmarkFtoa64FloatExp        12587         8263  -34.35%
strconv_test.BenchmarkFtoa64Big              7058         2825  -59.97%
json.BenchmarkCodeEncoder               357355200    276528200  -22.62%
json.BenchmarkCodeMarshal               360735200    279646400  -22.48%
json.BenchmarkCodeDecoder               731528600    709460600   -3.02%
json.BenchmarkCodeUnmarshal             754774400    731051200   -3.14%
json.BenchmarkCodeUnmarshalReuse        713379000    704218000   -1.28%
json.BenchmarkSkipValue                  51594300     51682600   +0.17%

benchmark                                old MB/s     new MB/s  speedup
json.BenchmarkCodeEncoder                    5.43         7.02    1.29x
json.BenchmarkCodeMarshal                    5.38         6.94    1.29x
json.BenchmarkCodeDecoder                    2.65         2.74    1.03x
json.BenchmarkCodeUnmarshal                  2.57         2.65    1.03x
json.BenchmarkCodeUnmarshalReuse             2.72         2.76    1.01x
json.BenchmarkSkipValue                     38.61        38.55    1.00x

R=golang-dev, r
CC=golang-dev
https://golang.org/cl/5369111
diff --git a/src/pkg/strconv/decimal.go b/src/pkg/strconv/decimal.go
index f572ea4..5415530 100644
--- a/src/pkg/strconv/decimal.go
+++ b/src/pkg/strconv/decimal.go
@@ -102,12 +102,6 @@
 	trim(a)
 }
 
-func newDecimal(i uint64) *decimal {
-	a := new(decimal)
-	a.Assign(i)
-	return a
-}
-
 // Maximum shift that we can do in one pass without overflow.
 // Signed int has 31 bits, and we have to be able to accommodate 9<<k.
 const maxShift = 27
@@ -303,32 +297,32 @@
 // If nd is zero, it means we're rounding
 // just to the left of the digits, as in
 // 0.09 -> 0.1.
-func (a *decimal) Round(nd int) *decimal {
+func (a *decimal) Round(nd int) {
 	if nd < 0 || nd >= a.nd {
-		return a
+		return
 	}
 	if shouldRoundUp(a, nd) {
-		return a.RoundUp(nd)
+		a.RoundUp(nd)
+	} else {
+		a.RoundDown(nd)
 	}
-	return a.RoundDown(nd)
 }
 
 // Round a down to nd digits (or fewer).
 // Returns receiver for convenience.
-func (a *decimal) RoundDown(nd int) *decimal {
+func (a *decimal) RoundDown(nd int) {
 	if nd < 0 || nd >= a.nd {
-		return a
+		return
 	}
 	a.nd = nd
 	trim(a)
-	return a
 }
 
 // Round a up to nd digits (or fewer).
 // Returns receiver for convenience.
-func (a *decimal) RoundUp(nd int) *decimal {
+func (a *decimal) RoundUp(nd int) {
 	if nd < 0 || nd >= a.nd {
-		return a
+		return
 	}
 
 	// round up
@@ -337,7 +331,7 @@
 		if c < '9' { // can stop after this digit
 			a.d[i]++
 			a.nd = i + 1
-			return a
+			return
 		}
 	}
 
@@ -346,7 +340,6 @@
 	a.d[0] = '1'
 	a.nd = 1
 	a.dp++
-	return a
 }
 
 // Extract integer part, rounded appropriately.
diff --git a/src/pkg/strconv/decimal_test.go b/src/pkg/strconv/decimal_test.go
index deb2e02..13a127f 100644
--- a/src/pkg/strconv/decimal_test.go
+++ b/src/pkg/strconv/decimal_test.go
@@ -70,17 +70,23 @@
 func TestDecimalRound(t *testing.T) {
 	for i := 0; i < len(roundtests); i++ {
 		test := &roundtests[i]
-		s := NewDecimal(test.i).RoundDown(test.nd).String()
+		d := NewDecimal(test.i)
+		d.RoundDown(test.nd)
+		s := d.String()
 		if s != test.down {
 			t.Errorf("Decimal %v RoundDown %d = %v, want %v",
 				test.i, test.nd, s, test.down)
 		}
-		s = NewDecimal(test.i).Round(test.nd).String()
+		d = NewDecimal(test.i)
+		d.Round(test.nd)
+		s = d.String()
 		if s != test.round {
 			t.Errorf("Decimal %v Round %d = %v, want %v",
 				test.i, test.nd, s, test.down)
 		}
-		s = NewDecimal(test.i).RoundUp(test.nd).String()
+		d = NewDecimal(test.i)
+		d.RoundUp(test.nd)
+		s = d.String()
 		if s != test.up {
 			t.Errorf("Decimal %v RoundUp %d = %v, want %v",
 				test.i, test.nd, s, test.up)
diff --git a/src/pkg/strconv/ftoa.go b/src/pkg/strconv/ftoa.go
index 07fe806..8342b6a 100644
--- a/src/pkg/strconv/ftoa.go
+++ b/src/pkg/strconv/ftoa.go
@@ -98,7 +98,8 @@
 	// The shift is exp - flt.mantbits because mant is a 1-bit integer
 	// followed by a flt.mantbits fraction, and we are treating it as
 	// a 1+flt.mantbits-bit integer.
-	d := newDecimal(mant)
+	d := new(decimal)
+	d.Assign(mant)
 	d.Shift(exp - int(flt.mantbits))
 
 	// Round appropriately.
@@ -184,7 +185,8 @@
 	// d = mant << (exp - mantbits)
 	// Next highest floating point number is mant+1 << exp-mantbits.
 	// Our upper bound is halfway inbetween, mant*2+1 << exp-mantbits-1.
-	upper := newDecimal(mant*2 + 1)
+	upper := new(decimal)
+	upper.Assign(mant*2 + 1)
 	upper.Shift(exp - int(flt.mantbits) - 1)
 
 	// d = mant << (exp - mantbits)
@@ -203,7 +205,8 @@
 		mantlo = mant*2 - 1
 		explo = exp - 1
 	}
-	lower := newDecimal(mantlo*2 + 1)
+	lower := new(decimal)
+	lower.Assign(mantlo*2 + 1)
 	lower.Shift(explo - int(flt.mantbits) - 1)
 
 	// The upper and lower bounds are possible outputs only if
diff --git a/src/pkg/strconv/internal_test.go b/src/pkg/strconv/internal_test.go
index 9a7f4f0..d0fa80e 100644
--- a/src/pkg/strconv/internal_test.go
+++ b/src/pkg/strconv/internal_test.go
@@ -6,7 +6,11 @@
 
 package strconv
 
-func NewDecimal(i uint64) *decimal { return newDecimal(i) }
+func NewDecimal(i uint64) *decimal {
+	d := new(decimal)
+	d.Assign(i)
+	return d
+}
 
 func SetOptimize(b bool) bool {
 	old := optimize