math/big: fix Float.Float32 conversion for denormal corner cases

The existing code was incorrect for numbers that after rounding would
become the smallest denormal float32 (instead the result was 0). This
caused all.bash to fail if Float32() were used in the compiler for
constant arithmetic (there's currently a work-around - see also issue
10321.

This change fixes the implementation of Float.Float32 and adds
corresponding test cases. Float32 and Float64 diverge at this point.
For ease of review, this change only fixes Float32. Float64 will be
made to match in a subsequent change.

Fixes #10321.

Change-Id: Iccafe37c1593a4946bc552e4ad2045f69be62d80
Reviewed-on: https://go-review.googlesource.com/10350
Reviewed-by: Alan Donovan <adonovan@google.com>

src/math/big/float.go

@@ -853,9 +853,6 @@ func (x *Float) Int64() (int64, Accuracy) {
 	panic("unreachable")
 }
 
-// TODO(gri) Float32 and Float64 are very similar internally but for the
-// floatxx parameters and some conversions. Should factor out shared code.
-
 // Float32 returns the float32 value nearest to x. If x is too small to be
 // represented by a float32 (|x| < math.SmallestNonzeroFloat32), the result
 // is (0, Below) or (-0, Above), respectively, depending on the sign of x.
@@ -880,64 +877,70 @@ func (x *Float) Float32() (float32, Accuracy) {
 			emax  = bias              //   127  largest unbiased exponent (normal)
 		)
 
-		// Float mantissae m have an explicit msb and are in the range 0.5 <= m < 1.0.
+		// Float mantissa m is 0.5 <= m < 1.0; compute exponent for floatxx mantissa.
-		// floatxx mantissae have an implicit msb and are in the range 1.0 <= m < 2.0.
+		e := x.exp - 1 // exponent for mantissa m with 1.0 <= m < 2.0
-		// For a given mantissa m, we need to add 1 to a floatxx exponent to get the
+		p := mbits + 1 // precision of normal float
-		// corresponding Float exponent.
-		// (see also implementation of math.Ldexp for similar code)
 
-		if x.exp < dmin+1 {
+		// If the exponent is too small, we may have a denormal number
-			// underflow
+		// in which case we have fewer mantissa bits available: reduce
-			if x.neg {
+		// precision accordingly.
-				var z float32
+		if e < emin {
-				return -z, Above
+			p -= emin - int(e)
+			// Make sure we have at least 1 bit so that we don't
+			// lose numbers rounded up to the smallest denormal.
+			if p < 1 {
+				p = 1
 			}
-			return 0.0, Below
 		}
-		// x.exp >= dmin+1
 
+		// round
 		var r Float
-		r.prec = mbits + 1 // +1 for implicit msb
+		r.prec = uint32(p)
-		if x.exp < emin+1 {
-			// denormal number - round to fewer bits
-			r.prec = uint32(x.exp - dmin)
-		}
 		r.Set(x)
+		e = r.exp - 1
 
 		// Rounding may have caused r to overflow to ±Inf
 		// (rounding never causes underflows to 0).
 		if r.form == inf {
-			r.exp = emax + 2 // cause overflow below
+			e = emax + 1 // cause overflow below
 		}
 
-		if r.exp > emax+1 {
+		// If the exponent is too large, overflow to ±Inf.
+		if e > emax {
 			// overflow
 			if x.neg {
 				return float32(math.Inf(-1)), Below
 			}
 			return float32(math.Inf(+1)), Above
 		}
-		// dmin+1 <= r.exp <= emax+1
 
-		var s uint32
+		// Determine sign, biased exponent, and mantissa.
-		if r.neg {
+		var sign, bexp, mant uint32
-			s = 1 << (fbits - 1)
+		if x.neg {
+			sign = 1 << (fbits - 1)
 		}
 
-		m := high32(r.mant) >> ebits & (1<<mbits - 1) // cut off msb (implicit 1 bit)
-
 		// Rounding may have caused a denormal number to
 		// become normal. Check again.
-		c := float32(1.0)
+		if e < emin {
-		if r.exp < emin+1 {
 			// denormal number
-			r.exp += mbits
+			if e < dmin {
-			c = 1.0 / (1 << mbits) // 2**-mbits
+				// underflow to ±0
+				if x.neg {
+					var z float32
+					return -z, Above
+				}
+				return 0.0, Below
+			}
+			// bexp = 0
+			mant = high32(r.mant) >> (fbits - r.prec)
+		} else {
+			// normal number: emin <= e <= emax
+			bexp = uint32(e+bias) << mbits
+			mant = high32(r.mant) >> ebits & (1<<mbits - 1) // cut off msb (implicit 1 bit)
 		}
-		// emin+1 <= r.exp <= emax+1
-		e := uint32(r.exp-emin) << mbits
 
-		return c * math.Float32frombits(s|e|m), r.acc
+		return math.Float32frombits(sign | bexp | mant), r.acc
 
 	case zero:
 		if x.neg {
@@ -956,6 +959,10 @@ func (x *Float) Float32() (float32, Accuracy) {
 	panic("unreachable")
 }
 
+// TODO(gri) Use same algorithm for Float64 as for Float32. The Float64 code
+// is incorrect for some corner cases (numbers that get rounded up to smallest
+// denormal - test cases are missing).
+
 // Float64 returns the float64 value nearest to x. If x is too small to be
 // represented by a float64 (|x| < math.SmallestNonzeroFloat64), the result
 // is (0, Below) or (-0, Above), respectively, depending on the sign of x.

src/math/big/float_test.go

@@ -203,6 +203,12 @@ func alike(x, y *Float) bool {
 	return x.Cmp(y) == 0 && x.Signbit() == y.Signbit()
 }
 
+func alike32(x, y float32) bool {
+	// we can ignore NaNs
+	return x == y && math.Signbit(float64(x)) == math.Signbit(float64(y))
+
+}
+
 func TestFloatMantExp(t *testing.T) {
 	for _, test := range []struct {
 		x    string
@@ -828,52 +834,69 @@ func TestFloatFloat32(t *testing.T) {
 		out float32
 		acc Accuracy
 	}{
-		{"-Inf", float32(math.Inf(-1)), Exact},
-		{"-0x1.ffffff0p2147483646", float32(-math.Inf(+1)), Below}, // overflow in rounding
-		{"-1e10000", float32(math.Inf(-1)), Below},                 // overflow
-		{"-0x1p128", float32(math.Inf(-1)), Below},                 // overflow
-		{"-0x1.ffffff0p127", float32(-math.Inf(+1)), Below},        // overflow
-		{"-0x1.fffffe8p127", -math.MaxFloat32, Above},
-		{"-0x1.fffffe0p127", -math.MaxFloat32, Exact},
-		{"-12345.000000000000000000001", -12345, Above},
-		{"-12345.0", -12345, Exact},
-		{"-1.000000000000000000001", -1, Above},
-		{"-1", -1, Exact},
-		{"-0x0.000002p-126", -math.SmallestNonzeroFloat32, Exact},
-		{"-0x0.000002p-127", -0, Above}, // underflow
-		{"-1e-1000", -0, Above},         // underflow
 		{"0", 0, Exact},
-		{"1e-1000", 0, Below},         // underflow
+
-		{"0x0.000002p-127", 0, Below}, // underflow
+		// underflow
-		{"0x0.000002p-126", math.SmallestNonzeroFloat32, Exact},
+		{"1e-1000", 0, Below},
+		{"0x0.000002p-127", 0, Below},
+		{"0x.0000010p-126", 0, Below},
+
+		// denormals
+		{"1.401298464e-45", math.SmallestNonzeroFloat32, Above}, // rounded up to smallest denormal
+		{"0x.ffffff8p-149", math.SmallestNonzeroFloat32, Above}, // rounded up to smallest denormal
+		{"0x.0000018p-126", math.SmallestNonzeroFloat32, Above}, // rounded up to smallest denormal
+		{"0x.0000020p-126", math.SmallestNonzeroFloat32, Exact},
+		{"0x.8p-148", math.SmallestNonzeroFloat32, Exact},
+		{"1p-149", math.SmallestNonzeroFloat32, Exact},
+		{"0x.fffffep-126", math.Float32frombits(0x7fffff), Exact}, // largest denormal
+
+		// normals
+		{"0x.ffffffp-126", math.Float32frombits(0x00800000), Above}, // rounded up to smallest normal
+		{"1p-126", math.Float32frombits(0x00800000), Exact},         // smallest normal
+		{"0x1.fffffep-126", math.Float32frombits(0x00ffffff), Exact},
+		{"0x1.ffffffp-126", math.Float32frombits(0x01000000), Above}, // rounded up
 		{"1", 1, Exact},
 		{"1.000000000000000000001", 1, Below},
 		{"12345.0", 12345, Exact},
 		{"12345.000000000000000000001", 12345, Below},
 		{"0x1.fffffe0p127", math.MaxFloat32, Exact},
 		{"0x1.fffffe8p127", math.MaxFloat32, Below},
-		{"0x1.ffffff0p127", float32(math.Inf(+1)), Above},        // overflow
+
-		{"0x1p128", float32(math.Inf(+1)), Above},                // overflow
+		// overflow
-		{"1e10000", float32(math.Inf(+1)), Above},                // overflow
+		{"0x1.ffffff0p127", float32(math.Inf(+1)), Above},
+		{"0x1p128", float32(math.Inf(+1)), Above},
+		{"1e10000", float32(math.Inf(+1)), Above},
 		{"0x1.ffffff0p2147483646", float32(math.Inf(+1)), Above}, // overflow in rounding
-		{"+Inf", float32(math.Inf(+1)), Exact},
+
+		// inf
+		{"Inf", float32(math.Inf(+1)), Exact},
 	} {
-		// conversion should match strconv where syntax is agreeable
+		for i := 0; i < 2; i++ {
-		if f, err := strconv.ParseFloat(test.x, 32); err == nil && float32(f) != test.out {
+			// test both signs
-			t.Errorf("%s: got %g; want %g (incorrect test data)", test.x, f, test.out)
+			tx, tout, tacc := test.x, test.out, test.acc
-		}
+			if i != 0 {
+				tx = "-" + tx
+				tout = -tout
+				tacc = -tacc
+			}
 
-		x := makeFloat(test.x)
+			// conversion should match strconv where syntax is agreeable
-		out, acc := x.Float32()
+			if f, err := strconv.ParseFloat(tx, 32); err == nil && !alike32(float32(f), tout) {
-		if out != test.out || acc != test.acc {
+				t.Errorf("%s: got %g; want %g (incorrect test data)", tx, f, tout)
-			t.Errorf("%s: got %g (%#x, %s); want %g (%#x, %s)", test.x, out, math.Float32bits(out), acc, test.out, math.Float32bits(test.out), test.acc)
+			}
-		}
 
-		// test that x.SetFloat64(float64(f)).Float32() == f
+			x := makeFloat(tx)
-		var x2 Float
+			out, acc := x.Float32()
-		out2, acc2 := x2.SetFloat64(float64(out)).Float32()
+			if !alike32(out, tout) || acc != tacc {
-		if out2 != out || acc2 != Exact {
+				t.Errorf("%s: got %g (%#x, %s); want %g (%#x, %s)", tx, out, math.Float32bits(out), acc, test.out, math.Float32bits(test.out), tacc)
-			t.Errorf("idempotency test: got %g (%s); want %g (Exact)", out2, acc2, out)
+			}
+
+			// test that x.SetFloat64(float64(f)).Float32() == f
+			var x2 Float
+			out2, acc2 := x2.SetFloat64(float64(out)).Float32()
+			if !alike32(out2, out) || acc2 != Exact {
+				t.Errorf("idempotency test: got %g (%s); want %g (Exact)", out2, acc2, out)
+			}
 		}
 	}
 }