JSON parsing of large 32-bit floating point values fails.

[jskeet]

The proto3 JSON spec doesn't give very much guidance on exactly how floating point values are represented.

The maximal finite 32-bit floating point value is precisely 340282346638528859811704183484516925440. Both .NET and (I believe) Go format this as 3.4028235e+38.

Unfortunately that fails to round-trip in .NET, because of the way we handle JSON parsing of 32-bit floating point numbers:

• Parse as a 64-bit float
• Compare with the maximal 32-bit float - if it's less than or equal, proceed, otherwise throw an exception

There are 64-bit floating point values closer to 3.4028235e+38 than the max 32-bit value, so we end up throwing an exception.
We should, instead, try casting to float and see whether the result is infinite.

[jskeet]

@mcy: This feels like something that could usefully be in the JSON conformance test suite. (I'm adding unit tests to the C# code of course.)

[haberman]

The conformance tests already have a "max float" case, but they are testing for 3.402823e+38 rather than 3.4028235e+38:

protobuf/conformance/binary_json_conformance_suite.cc

Lines 2085 to 2088 in f8be281

	RunValidJsonTest(
	"FloatFieldMaxPositiveValue", REQUIRED,
	R"({"optionalFloat": 3.402823e+38})",
	"optional_float: 3.402823e+38");

Also it appears the min float case has a copy-and-paste bug, where it's accidentally checking the max float value again:

protobuf/conformance/binary_json_conformance_suite.cc

Lines 2089 to 2092 in f8be281

	RunValidJsonTest(
	"FloatFieldMinNegativeValue", REQUIRED,
	R"({"optionalFloat": 3.402823e+38})",
	"optional_float: 3.402823e+38");

[jskeet]

@haberman: Would you like me to make changes in that file in #10514, or should we keep that separate?