Fix issue 15885 - float serialized to JSON loses precision

[tsbockman]

This is a continuation of @bbasile 's PR #4343 .

[dlang-bot]

Fix	Bugzilla	Description
✓	15885	float serialized to JSON loses precision

[ghost]

What 's happening on Windows ? If the failure is related to DigitalMars snn.lib then the version should be changed to CRuntime_DigitalMars because MS C standard lib might not have the issue.

[tsbockman]

What 's happening on Windows ?

I don't know. Win_32 failed on win-farm-2at assert(test(1.223e+24)), so I added the debug printout and re-ran the test... which then passed on win-farm-1.

Originally I thought it was just imprecision in the Windows snprintf() - which is indirectly called by toJSON() - but now I'm wondering if it's actually a code gen bug, or maybe flaky hardware on one/both of the Windows test servers.

[tsbockman]

@bbasile I think there is a code gen bug on Win_32:

core.exception.AssertError@std\json.d(1740):
    "1.22300000000000003e+24": 1.22300000000000003e+24 != 1.22300000000000003e+24

And from another run, printing the bit pattern instead:

core.exception.AssertError@std\json.d(1740):
    "1.22300000000000003e+24": 44F02FAF67D2C4A9 != 44F02FAF67D2C4A9

Reducing this could be a pain though, since I don't have a Win_32 development environment of my own right now.

[ghost]

We have missed this: https://issues.dlang.org/show_bug.cgi?id=15861. The fix should be OK now !

[tsbockman]

@bbasile

The fix should be OK now !

You are correct. Excellent!

It's now 100% precise on all platforms except Win32, where the result is sometimes off by a single bit, which is acceptable for floating-point. (Or, at least a vast improvement over the current bugged behaviour.)

[tsbockman]

Ping @klickverbot @CyberShadow . This important JSON bug fix passes the auto-tester now; does anyone want to merge it?

[andralex]

Auto-merge toggled on