Implement (some) floating point comparisons.

vext01 commented 2 weeks ago

Float comparisons are a bit more involved than integer comparisons.

Here's some of the highlights (and my working understanding of them):

There are some special values that make the ordering a bit more interesting, e.g. +inf, -inf, NaN, +0 and -0.
There is a large encoding space for representing NaNs and NaNs are, in float terminology, "unordered". i.e. in the partial order of floats a NaN cannot be compared to any other element. Amusingly, two NaNs are never considered equal.
This plays into comparison operations. You can get the unordered result if at least one of the things you want to compare are NaN.
NaN comes in two variants: quiet and signalling, which basically mean "silently proceed, despite NaN" or "crash". In our x86 codegen we use ucomis{s,d} and comis{s,d} respectively for this. The latter would cause the chip to raise a floating point exception.
But what happens when we get the unordered result for a comparison and we are not crashing? We still need an answer, were they equal or not? We do as clang does, and follow the IEE-754 float spec:
- Any "not equal" comparison involving NaN is true.
- All other comparisons are false.

I hope I've got this all right :)

vext01 commented 2 weeks ago

(After this attrib.lua gets a little further)

vext01 commented 2 weeks ago

I'll force push to fix the conflicts, if that's ok?

vext01 commented 2 weeks ago

vext01 commented 2 weeks ago

OK to force push a squash/rebase?

ltratt commented 2 weeks ago

Please go ahead.

vext01 commented 2 weeks ago

ready

vext01 commented 2 weeks ago

force pushed a fix for formatting. go ahead.

ykjit / yk