Fix JS number rounding on x87

[bnoordhuis]

Disable 80-bits extended precision, use standard 64-bits precision. The extra precision affects rounding and is user-observable, meaning it causes test262 tests to fail.

The rounding mode is a per-CPU (or, more accurately, per-FPU) property, and since quickjs is often used as a library, take care to tweak the FPU control word before sensitive operations, and restore it afterwards.

Only affects x87. SSE is IEEE-754 conforming.

Fixes: #1236

(new CI job can fail, may need to link explicitly to libm)

[saghul]

I'm not going to pretend to understand what's going on here 😅

[bnoordhuis]

/home/runner/work/_temp/d0c8a954-89f3-4555-9dc2-6e488b6f3ac2: line 1: git: not found

Which I guess makes sense because git isn't part of build-base but then how the heck do other jobs pass? Anyway, I'll add it and let's see how it goes.

[bnoordhuis]

I'm not going to pretend to understand what's going on here 😅

The x87 coprocessor has a register that controls various options, like rounding, exceptions, etc. Its floating-point registers are 80 bits wide, which is good if you want extra precision, but bad if you want strict IEEE-754 conformance.

I made it turn off the extra precision while we add/subtract/etc., then restore the coprocessor to its original state, so we don't affect other parts of the (embedder's) program, that may depend on that extra precision.

[saghul]

I'm not going to pretend to understand what's going on here 😅

The x87 coprocessor has a register that controls various options, like rounding, exceptions, etc. Its floating-point registers are 80 bits wide, which is good if you want extra precision, but bad if you want strict IEEE-754 conformance.

I made it turn off the extra precision while we add/subtract/etc., then restore the coprocessor to its original state, so we don't affect other parts of the (embedder's) program, that may depend on that extra precision.

Today in "learning stuff with Ben"! Thank you, TIL!