f80 float round-trip failure on windows x86_64

[tiehuis]

Zig Version

0.14.0-dev.283+1d20ff11d

Steps to Reproduce and Observed Behavior

When adding f80 support (#20462) there were CI failures on the windows x86_64 build when performing round-trip checks on f80. This specific case has been disabled. Other platforms correctly round-trip these.

An example of a specific failing test case:

C:\Users\ci\george2\_work\zig\zig\build-release\zig-local-cache\o\ef0547664f7b4931546bf402c80a8059\test.exe --listen=- 
test
+- test-std
   +- run test std-x86_64-windows.win11_ge...win11_ge-gnu-skylake-Debug 2657/2724 passed, 1 failed, 66 skipped
error: 'fmt.format_float.test.format f80' failed: expected 907682844368439810595840, found 907682844368439810596864
C:\Users\ci\george2\_work\zig\zig\lib\std\testing.zig:93:17: 0xe829e3 in expectEqualInner__anon_35380 (test.exe.obj)
                return error.TestExpectedEqual;
                ^
C:\Users\ci\george2\_work\zig\zig\lib\std\fmt\format_float.zig:1530:9: 0x12db79a in check__anon_46210 (test.exe.obj)
        try std.testing.expectEqual(value_bits, o_bits);
        ^
C:\Users\ci\george2\_work\zig\zig\lib\std\fmt\format_float.zig:1655:5: 0x12dcd86 in test.format f80 (test.exe.obj)
    try check(f80, -2.109808898695963e16, "-2.109808898695963e16");
    ^

input = -2.109808898695963e16

i = 907682844368439810595840 // in
r = 907682844368439810596864 // reparsed

i = 0xc03595e9334703be1c00
r = 0xc03595e9334703be2000

i_man = 0x95e9334703be1c00
i_exp = 0x4035 (16437)
i_int = 1
i_sgn = 1
 
r_man = 0x95e9334703be2000
r_exp = 0x4035 (16437)
r_int = 1
r_sgn = 1

Note that r_man - i_man = 1024 = 2^8 which seems a little suspect.

Expected Behavior

We should be able to round-trip all f80's:

zig/lib/std/fmt/format_float.zig

Line 1526 in 944c6d4

if (T == f80 and builtin.target.os.tag == .windows and builtin.target.cpu.arch == .x86_64) return;