reflect: when Converting between float32s, don't lose signal NaNs

Trying this CL again, with a test that skips 387.

When converting from float32->float64->float32, any signal NaNs
get converted to quiet NaNs. Avoid that so using reflect.Value.Convert
between two float32 types keeps the signal bit of NaNs.

Skip the test on 387. I don't see any sane way of ensuring that a
float load + float store is faithful on that platform.

Fixes #36400

Change-Id: Ic316c74ddc155632e40424e207375b5d50dcd853
Reviewed-on: https://go-review.googlesource.com/c/go/+/221792
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>

2 files changed, 45 insertions, 0 deletions

diff --git a/src/reflect/all_test.go b/src/reflect/all_test.go
index 00c18104eb..66d9661aeb 100644
--- a/src/reflect/all_test.go
+++ b/src/reflect/all_test.go
@@ -4163,6 +4163,37 @@ func TestConvert(t *testing.T) {
 	}
 }
 
+var gFloat32 float32
+
+func TestConvertNaNs(t *testing.T) {
+	const snan uint32 = 0x7f800001
+
+	// Test to see if a store followed by a load of a signaling NaN
+	// maintains the signaling bit. The only platform known to fail
+	// this test is 386,GO386=387. The real test below will always fail
+	// if the platform can't even store+load a float without mucking
+	// with the bits.
+	gFloat32 = math.Float32frombits(snan)
+	runtime.Gosched() // make sure we don't optimize the store/load away
+	r := math.Float32bits(gFloat32)
+	if r != snan {
+		// This should only happen on 386,GO386=387. We have no way to
+		// test for 387, so we just make sure we're at least on 386.
+		if runtime.GOARCH != "386" {
+			t.Errorf("store/load of sNaN not faithful")
+		}
+		t.Skip("skipping test, float store+load not faithful")
+	}
+
+	type myFloat32 float32
+	x := V(myFloat32(math.Float32frombits(snan)))
+	y := x.Convert(TypeOf(float32(0)))
+	z := y.Interface().(float32)
+	if got := math.Float32bits(z); got != snan {
+		t.Errorf("signaling nan conversion got %x, want %x", got, snan)
+	}
+}
+
 type ComparableStruct struct {
 	X int
 }
diff --git a/src/reflect/value.go b/src/reflect/value.go
index 51e7d195fe..08f0d259de 100644
--- a/src/reflect/value.go
+++ b/src/reflect/value.go
@@ -2541,6 +2541,14 @@ func makeFloat(f flag, v float64, t Type) Value {
 	return Value{typ, ptr, f | flagIndir | flag(typ.Kind())}
 }
 
+// makeFloat returns a Value of type t equal to v, where t is a float32 type.
+func makeFloat32(f flag, v float32, t Type) Value {
+	typ := t.common()
+	ptr := unsafe_New(typ)
+	*(*float32)(ptr) = v
+	return Value{typ, ptr, f | flagIndir | flag(typ.Kind())}
+}
+
 // makeComplex returns a Value of type t equal to v (possibly truncated to complex64),
 // where t is a complex64 or complex128 type.
 func makeComplex(f flag, v complex128, t Type) Value {
@@ -2613,6 +2621,12 @@ func cvtUintFloat(v Value, t Type) Value {
 
 // convertOp: floatXX -> floatXX
 func cvtFloat(v Value, t Type) Value {
+	if v.Type().Kind() == Float32 && t.Kind() == Float32 {
+		// Don't do any conversion if both types have underlying type float32.
+		// This avoids converting to float64 and back, which will
+		// convert a signaling NaN to a quiet NaN. See issue 36400.
+		return makeFloat32(v.flag.ro(), *(*float32)(v.ptr), t)
+	}
 	return makeFloat(v.flag.ro(), v.Float(), t)
 }