Fix issue 16432 - JSON incorrectly parses to string

[wilzbach]

The entire json module seems to require a bit of love - though I am still hoping for std.data.json

[dlang-bot]

Fix	Bugzilla	Description
✓	16432	JSON incorrectly parses to string

[wilzbach]

Great it seems that the floating point behavior on Win32_64 is different to all other machines and architectures (yep - even Win32). This could be due to a different C runtime which provides snprintf which is used by formattedWrite in std.format.
If someone has a 64-bit Windows machine, it would be great to know what the following yields:

import std.format, std.stdio;
format("%.18g", 1e18 - 65).writeln;
format("%.18g", 1e18 - 64).writeln;

[WebDrake]

@wilzbach why not just add in a temporary patch which does the writelns in the unittests, then check the auto-tester's console output? Or include the formatted string in the assert error message?

[ghost]

This is why Grisu or RYU would be needed. These small differences due to different standard C library printf would be fixed for good.

[berni44]

Great it seems that the floating point behavior on Win32_64 is different to all other machines and architectures (yep - even Win32). This could be due to a different C runtime which provides snprintf which is used by formattedWrite in std.format.
If someone has a 64-bit Windows machine, it would be great to know what the following yields:

import std.format, std.stdio;
format("%.18g", 1e18 - 65).writeln;
format("%.18g", 1e18 - 64).writeln;

Meanwhile snprintf is removed for %g (at least for double and float). Unfortunately this is exactly the gap between %e-like and %f-like output, which may still differ between different computers. The reason for this is the way D handles floating point math (internal promotion to real) and the fact, that floating point math might differ slightly, depending on the way the calculations are optimized by the compiler. To avoid this, a software floating point unit with predictable results would be needed, IMHO an overkill. See #7836 for more on this.

[RazvanN7]

Closing in favor of the newer PR.