Repository navigation
Conversation
…rmat DecimalCodec.EncodeDecimal wrote coefficients with more digits than the format allows, and exponents outside its range, without any check (only a Debug.Assert). In Release builds the excess bits overflowed into the combination/exponent fields and the server stored a different value, e.g.: 12345678901234567890123456789012345 -> DECFLOAT(34) stored 8345678901234567890123456789012345 30000000000000004E-17 -> DECFLOAT(16) stored 0.08000000000000004 1E400 -> DECFLOAT(16) stored 8.000000000000001E-353 The encoder now brings the value into the format's ranges without changing its numeric value (dropping trailing zeros of the coefficient, folding the exponent into the coefficient, clamping the exponent of zero), and throws OverflowException when the value cannot be represented exactly (too many significant digits, exponent overflow or underflow). Values that previously got corrupted although representable (e.g. 1E384 in DECFLOAT(16)) now work.
The conversions formatted the value with a fixed custom format and then
removed the decimal point, assuming one was present. This threw
ArgumentOutOfRangeException for every value without a fractional part,
e.g. (FbDecFloat)1m, (FbDecFloat)100m, (FbDecFloat)2.0, and for values
smaller than the format's decimal places, e.g. (FbDecFloat)1e-20. It also
dropped the scale of decimals (1.10m became 1.1) and digits beyond 15
decimal places of doubles (7 for floats).
- decimal: converted exactly from its 96-bit integer and scale, so the
value and its scale are preserved (1.10m -> 110E-2), including -0.
- double/float: converted to the shortest decimal that round-trips to the
same binary value ("R" format), e.g. 0.1 -> 1E-1, 0.1 + 0.2 ->
30000000000000004E-17, 1e-20 -> 1E-20; -0.0 -> negative zero.
A converted double may have up to 17 significant digits; sending it to a
DECFLOAT(16) throws OverflowException instead of being silently rounded
(this relies on the encoder range check of the previous commit).
Random.NextInt64() never returns a negative value, so the random doubles were all positive. Use the full long range and assert both signs are covered.
Parameters: a DECFLOAT parameter only accepted FbDecFloat; any other value (decimal, double, int, ...) threw InvalidCastException, even when the parameter was created by AddWithValue with a decimal. Values of type decimal, double, float, the integral types and BigInteger are now converted to FbDecFloat (decimal exactly, double/float to the shortest round-trip value). A value that does not fit the DECFLOAT precision exactly throws OverflowException instead of being rounded. Reader: GetDecimal, GetDouble, GetFloat, GetInt16/32/64 and GetByte (and the matching GetFieldValue<T>) on a DECFLOAT column threw InvalidCastException because FbDecFloat is not IConvertible. Now: - GetDecimal: exact, keeping the scale (1.10 -> 1.10m); OverflowException when the value cannot be represented exactly (too large, too many digits, NaN, infinity). - GetDouble/GetFloat: nearest value, like on NUMERIC columns; NaN and infinities are mapped, out-of-range values become infinity or zero. - Integer getters: nearest integer, ties to even, as Convert does for decimal columns; OverflowException when out of range, NaN or infinity. FbDecFloat gets the matching public explicit conversions to decimal, double and float.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #1300. Based on #1306.
Summary
DECFLOAT columns and parameters now work with the usual .NET numeric types. Values are never rounded silently: a value that does not fit exactly throws
OverflowException.Changes
Parameters (
DbValue.GetDecFloat):decimal,double,float, the integral types andBigIntegerare converted toFbDecFloat, using the rules of #1306. They used to throwInvalidCastException, also withAddWithValue.Reader (
DbValuegetters, soGetFieldValue<T>too):GetDecimal1.10becomes1.10m);OverflowExceptionwhen the value is not exact (too large, too many digits, NaN or infinity)GetDouble,GetFloatGetInt16,GetInt32,GetInt64,GetByteConvertdoes fordecimalcolumns);OverflowExceptionwhen out of range, NaN or infinityFbDecFloatgets the matching public explicit conversions todecimal,doubleandfloat.docs/decfloat.mddocuments the rules.Writing is stricter than the server's own
CAST. Adoublewhose shortest exact form has 17 digits (e.g.0.1 + 0.2) throws for aDECFLOAT(16)parameter, whereasCAST(0.1e0 + 0.2e0 AS DECFLOAT(16))rounds to0.3000000000000000. This is deliberate (see #1306): the provider does not round a value silently. Values that fit exactly are stored unchanged, as the round-trip tests below show.Tests
FbDecFloatTypeTests(NoServer): the new conversions, including limits, NaN, infinities, negative zero and rounding to even.FbDecFloat16SupportTestsandFbDecFloat34SupportTests:GetDecimal,GetDouble,GetFloatandGetInt*: exact, nearest, rounding and throwing cases.DbValuechanges, 92 of the new tests fail. With them, the embedded + NoServer suite passes: 834 passed, 0 failed.