sql/parser: avoid float -> unsigned conversions

While investigating this, I discovered that our current `round`
implementation is quite dubious, and does not rely on any referenced
source material or any material that I could find. I also found that
Postgres does not implement 2-ary `round` where the first argument is
a float.

The above is addressed by:
- replacing `round(float)` with a transcription of Postgres' `rint`.
- replacing `round(float, int)` with an implementation that round-trips
  through apd. This is likely much slower, but likely correct.

Updates #14405.

Author: tamird

pkg/cmd/metacheck/main.go

@@ -17,6 +17,8 @@
 package main
 
 import (
+	"go/ast"
+	"go/types"
 	"log"
 	"os"
 
@@ -38,7 +40,9 @@ func (m *metaChecker) Init(program *lint.Program) {
 }
 
 func (m *metaChecker) Funcs() map[string]lint.Func {
-	funcs := make(map[string]lint.Func)
+	funcs := map[string]lint.Func{
+		"FloatToUnsigned": checkConvertFloatToUnsigned,
+	}
 	for _, checker := range m.checkers {
 		for k, v := range checker.Funcs() {
 			if _, ok := funcs[k]; ok {
@@ -51,6 +55,57 @@ func (m *metaChecker) Funcs() map[string]lint.Func {
 	return funcs
 }
 
+// @ianlancetaylor via golang-nuts[0]:
+//
+// For the record, the spec says, in https://golang.org/ref/spec#Conversions:
+// "In all non-constant conversions involving floating-point or complex
+// values, if the result type cannot represent the value the conversion
+// succeeds but the result value is implementation-dependent."  That is the
+// case that applies here: you are converting a negative floating point number
+// to uint64, which can not represent a negative value, so the result is
+// implementation-dependent.  The conversion to int64 works, of course. And
+// the conversion to int64 and then to uint64 succeeds in converting to int64,
+// and when converting to uint64 follows a different rule: "When converting
+// between integer types, if the value is a signed integer, it is sign
+// extended to implicit infinite precision; otherwise it is zero extended. It
+// is then truncated to fit in the result type's size."
+//
+// So, basically, don't convert a negative floating point number to an
+// unsigned integer type.
+//
+// [0] https://groups.google.com/d/msg/golang-nuts/LH2AO1GAIZE/PyygYRwLAwAJ
+//
+// TODO(tamird): upstream this.
+func checkConvertFloatToUnsigned(j *lint.Job) {
+	fn := func(node ast.Node) bool {
+		call, ok := node.(*ast.CallExpr)
+		if !ok {
+			return true
+		}
+		castType, ok := j.Program.Info.TypeOf(call.Fun).(*types.Basic)
+		if !ok {
+			return true
+		}
+		if castType.Info()&types.IsUnsigned == 0 {
+			return true
+		}
+		for _, arg := range call.Args {
+			argType, ok := j.Program.Info.TypeOf(arg).(*types.Basic)
+			if !ok {
+				continue
+			}
+			if argType.Info()&types.IsFloat == 0 {
+				continue
+			}
+			j.Errorf(arg, "do not convert a floating point number to an unsigned integer type")
+		}
+		return true
+	}
+	for _, f := range j.Program.Files {
+		ast.Inspect(f, fn)
+	}
+}
+
 func main() {
 	unusedChecker := unused.NewChecker(unused.CheckAll)
 	unusedChecker.WholeProgram = true

pkg/sql/parser/builtins.go

@@ -1263,7 +1263,7 @@ var Builtins = map[string][]Builtin{
 
 	"round": {
 		floatBuiltin1(func(x float64) (Datum, error) {
-			return round(x, 0)
+			return NewDFloat(DFloat(round(x))), nil
 		}, "Rounds `val` to the nearest integer using half to even (banker's) rounding."),
 		decimalBuiltin1(func(x *apd.Decimal) (Datum, error) {
 			return roundDecimal(x, 0)
@@ -1273,7 +1273,25 @@ var Builtins = map[string][]Builtin{
 			Types:      ArgTypes{{"input", TypeFloat}, {"decimal_accuracy", TypeInt}},
 			ReturnType: fixedReturnType(TypeFloat),
 			fn: func(_ *EvalContext, args Datums) (Datum, error) {
-				return round(float64(*args[0].(*DFloat)), int64(MustBeDInt(args[1])))
+				var x apd.Decimal
+				if _, err := x.SetFloat64(float64(*args[0].(*DFloat))); err != nil {
+					return nil, err
+				}
+
+				// TODO(mjibson): make sure this fits in an int32.
+				scale := int32(MustBeDInt(args[1]))
+
+				var d apd.Decimal
+				if _, err := RoundCtx.Quantize(&d, &x, -scale); err != nil {
+					return nil, err
+				}
+
+				f, err := d.Float64()
+				if err != nil {
+					return nil, err
+				}
+
+				return NewDFloat(DFloat(f)), nil
 			},
 			Info: "Keeps `decimal_accuracy` number of figures to the right of the zero position " +
 				" in `input` using half to even (banker's) rounding.",
@@ -2082,36 +2100,72 @@ func overlay(s, to string, pos, size int) (Datum, error) {
 	return NewDString(string(runes[:pos]) + to + string(runes[after:])), nil
 }
 
-func round(x float64, n int64) (Datum, error) {
-	pow := math.Pow(10, float64(n))
+// Transcribed from Postgres' src/port/rint.c, with c-style comments preserved
+// for ease of mapping.
+//
+// https://github.com/postgres/postgres/blob/REL9_6_3/src/port/rint.c
+func round(x float64) float64 {
+	/* Per POSIX, NaNs must be returned unchanged. */
+	if math.IsNaN(x) {
+		return x
+	}
 
-	if pow == 0 {
-		// Rounding to so many digits on the left that we're underflowing.
-		// Avoid a NaN below.
-		return NewDFloat(DFloat(0)), nil
+	/* Both positive and negative zero should be returned unchanged. */
+	if x == 0.0 {
+		return x
 	}
-	if math.Abs(x*pow) > 1e17 {
-		// Rounding touches decimals below float precision; the operation
-		// is a no-op.
-		return NewDFloat(DFloat(x)), nil
+
+	roundFn := math.Ceil
+	if math.Signbit(x) {
+		roundFn = math.Floor
 	}
 
-	v, frac := math.Modf(x * pow)
-	// The following computation implements unbiased rounding, also
-	// called bankers' rounding. It ensures that values that fall
-	// exactly between two integers get equal chance to be rounded up or
-	// down.
-	if x > 0.0 {
-		if frac > 0.5 || (frac == 0.5 && uint64(v)%2 != 0) {
-			v += 1.0
-		}
-	} else {
-		if frac < -0.5 || (frac == -0.5 && uint64(v)%2 != 0) {
-			v -= 1.0
-		}
+	/*
+	 * Subtracting 0.5 from a number very close to -0.5 can round to
+	 * exactly -1.0, producing incorrect results, so we take the opposite
+	 * approach: add 0.5 to the negative number, so that it goes closer to
+	 * zero (or at most to +0.5, which is dealt with next), avoiding the
+	 * precision issue.
+	 */
+	xOrig := x
+	x -= math.Copysign(0.5, x)
+
+	/*
+	 * Be careful to return minus zero when input+0.5 >= 0, as that's what
+	 * rint() should return with negative input.
+	 */
+	if x == 0 || math.Signbit(x) != math.Signbit(xOrig) {
+		return math.Copysign(0.0, xOrig)
+	}
+
+	/*
+	 * For very big numbers the input may have no decimals.  That case is
+	 * detected by testing x+0.5 == x+1.0; if that happens, the input is
+	 * returned unchanged.  This also covers the case of minus infinity.
+	 */
+	if x == xOrig-math.Copysign(1.0, x) {
+		return xOrig
+	}
+
+	/* Otherwise produce a rounded estimate. */
+	r := roundFn(x)
+
+	/*
+	 * If the rounding did not produce exactly input+0.5 then we're done.
+	 */
+	if r != x {
+		return r
 	}
 
-	return NewDFloat(DFloat(v / pow)), nil
+	/*
+	 * The original fractional part was exactly 0.5 (since
+	 * floor(input+0.5) == input+0.5).  We need to round to nearest even.
+	 * Dividing input+0.5 by 2, taking the floor and multiplying by 2
+	 * yields the closest even number.  This part assumes that division by
+	 * 2 is exact, which should be OK because underflow is impossible
+	 * here: x is an integer.
+	 */
+	return roundFn(x*0.5) * 2.0
 }
 
 func roundDecimal(x *apd.Decimal, n int32) (Datum, error) {

pkg/sql/parser/decimal.go

@@ -32,4 +32,11 @@ var (
 	ExactCtx = DecimalCtx.WithPrecision(0)
 	// HighPrecisionCtx is a decimal context with high precision.
 	HighPrecisionCtx = DecimalCtx.WithPrecision(2000)
+	// RoundCtx is a decimal context with high precision and RoundHalfEven
+	// rounding.
+	RoundCtx = func() *apd.Context {
+		ctx := *HighPrecisionCtx
+		ctx.Rounding = apd.RoundHalfEven
+		return &ctx
+	}()
 )

pkg/sql/testdata/logic_test/builtin_function

@@ -655,6 +655,9 @@ SELECT radians(-45.0), radians(45.0)
 ----
 -0.7853981633974483 0.7853981633974483
 
+query error invalid operation
+SELECT round(123.456::float, -2438602134409251682)
+
 query RRR
 SELECT round(4.2::float, 0), round(4.2::float, 10), round(4.22222222::decimal, 3)
 ----
@@ -710,12 +713,12 @@ SELECT round(-1.7976931348623157e+308::float, 1), round(1.7976931348623157e+308:
 query RR
 SELECT round(-1.7976931348623157e+308::float, -303), round(1.7976931348623157e+308::float, -303)
 ----
--1.797690000000001e+308 1.797690000000001e+308
+-1.79769e+308 1.79769e+308
 
 query RR
 SELECT round(-1.23456789e+308::float, -308), round(1.23456789e+308::float, -308)
 ----
--1.0000000000000006e+308 1.0000000000000006e+308
+-1e+308 1e+308
 
 query RR
 SELECT round(-1.7976931348623157e-308::float, 1), round(1.7976931348623157e-308::float, 1)
@@ -727,10 +730,10 @@ SELECT 1.234567890123456789::float,  round(1.234567890123456789::float,15), roun
 ----
 1.2345678901234567 1.234567890123457 1.2345678901234567 1.2345678901234567
 
-query RRRR
-SELECT round(123.456::float, -1), round(123.456::float, -2), round(123.456::float, -3), round(123.456::float, -2438602134409251682)
+query RRR
+SELECT round(123.456::float, -1), round(123.456::float, -2), round(123.456::float, -3)
 ----
-120 100 0 0
+120 100 0
 
 query RRRR
 SELECT round(123.456::decimal, -1), round(123.456::decimal, -2), round(123.456::decimal, -3), round(123.456::decimal, -200)