GitLab

The bug:
DecimalFormat.format() behavior is slightly
lossy around 15 decimal digits even without any
digit constraints.

This change isolates the test failures that result from
this bug to 2 test cases:

test_formatDouble_bug17656132
test_formatDouble_roundingProblemCases

Example of the bug:

Double: 999999999999999.9 is represented as an IEEE 754
value which is exactly decimal 999999999999999.875

When format(999999999999999.9) is called on a DecimalFormat
with large MaxIntegerDigit and MaxFractionDigit....

Correct answer: "999999999999999.9"
Actual answer: "1000000000000000"

By contrast Double.toString() prints 9.999999999999999E14
for Android and the RI (correctly).

The DecimalFormat is printing to 16 decimal digits: The
inclusion of the 16th digit implies slightly more precision
than IEEE 754 provides (~15.9 decimal digits for most of the
representable range).
However, the use of 16 decimal digits for outputting IEEE 754
appears consistent with Double.toString() and elsewhere.

Before printing, DecimalFormat appears to be rounding to
15 decimal digits internally (or something similar).

Parsing "1000000000000000" produces a different double
representation than the original double
(one that is closer to 1000000000000000 than
999999999999999.9). This is the bug - we just lost information.
We should be able to round-trip a double if there is no rounding
since every double is representable with decimal and we have
sufficient digits available to represent it (close enough) in
decimal.

Additional tests have been added to demonstrate the bug
and also demonstrate the (correct) formatting behavior when the
formatter is rounding.

test_formatDouble_maxFractionDigits: rounding at 1, 10, 14 digits
after the dp.
test_formatDouble_roundingTo15Digits: rounding at 15 total digits
overall
test_formatDouble_bug17656132: Demonstrates the bug concisely.

The test changes:

test_formatDouble_wideRange():
implicitly assumed that the any loss of accuracy
when a decimal string was turned into a double
would be undone when format() was called, and it would
always arrive back at the original string. The test
has been re-written here to use BigDecimal rather than
Double.parseDouble(), and to compare two doubles rather
than original string with the output from format().

The test was previously failing with the RI for 1E23:
the closest representable double to 1E23 is exactly
99999999999999991611392.
The value produces "99999999999999990000000" when formatted
with the RI and not "100000000000000000000000". On Android
it was passing for 1E23 because of bug 17656132 rounding
back to the original decimal value.

This test was previously failing on Android with 1E-309 because
below 1E-308 IEEE 754 starts losing precision and the closest
representable value is not close to the original string. The
test now isn't affected if the double being tested is not close
to the original decimal; it passes providing the can be round
tripped.

test_formatDouble_roundingProblemCases: Re-written like the
_wideRange test but continues to demonstrate the bug due to
the test values (intentionally) chosen.

Bug: 17656132
Change-Id: I7d81e38bd1f9dbfd1e1b2caa60a6bb16b871b925