i686 test262 failures

[bnoordhuis]

gcc 14.2.0

$ ./build/x86/run-test262 -c test262.conf
test262/test/built-ins/Date/UTC/fp-evaluation-order.js:19: unexpected error: Test262Error: order of operations / precision in MakeTime Expected SameValue(«29256», «29312») to be true
test262/test/language/types/number/S8.5_A2.1.js:16: unexpected error: Test262Error: #1: var x = 9007199254740994.0; var y = 1.0 - 1/65536.0; var z = x + y; var d = z - x; d === 0. Actual: 2
test262/test/language/types/number/S8.5_A2.2.js:16: unexpected error: Test262Error: #1: var x = 9007199254740994.0; var y = 1.0 - 1/65536.0; var z = x + y; var d = z - x; d !== 2
Result: 57/41861 errors, 3285 excluded, 5569 skipped, 3 new

All three seem to be related to floating-point rounding. Tested with -fno-fast-math and volatile-ing all the things - no change.

[bavier]

d would end up as 2 if z were evaluated as z = (x + 1) - 1/65536.0 instead of the intended z = x + (1 - 1/65536.0).

[bnoordhuis]

It's because the x87 FPU rounds intermediate results with 80 bits of precision instead of 64 bits. It's possible to replicate on x86_64 with -mno-sse. It's strictly speaking not even a quickjs bug; any C program that tries that exact calculation is affected.

Anyway, long story short, tweaking the FPU control word fixes it: #1244