Use case for endian-specific floats f16le, f16be etc

[gitlost]

Was wondering what the use case for endian-specific floats was, as currently they're broken (for f16 and f32 anyway). E.g. on a little-endian machine, the following

package check_float_endian

import "core:fmt"

main :: proc() {
	f :f16be = 1

	fmt.printf("f16be %v (%h)\n", f, f);

	f2 :f32be = 1

	fmt.printf("f32be %v (%h)\n", f2, f2);
}

will print

f16be 0.000 (0h0)
f32be 0.000 (0h0)

Of course this can be fixed but if they're not being used much (which them being broken sort of hints at) then removing them from the language would be a very nice simplification.

[Kelimion]

This might be a bug in fmt?
Does fmt.printf("0x%04x\n", transmute(u32)f2) look like it should for that value as a big endian float?

We had a problem with u32be initially as well, but it's exercised as part of the PNG test suite decoding the gAMA chunk and testing the value, so that got fixed.

[gitlost]

Hi, it might be, but was looking at the case ExactValue_Float: in lb_const_value (https://github.com/odin-lang/Odin/blob/master/src/llvm_backend_const.cpp#L579) and noticed it just does a swap of the f64 value (as opposed to the case ExactValue_Complex following which calls lb_const_f16 etc which swap based on endianness) so I don't think so.

I noticed the use of endian-specific ints alright but the float ones seem much less used. There's also the issue of what endianness means for floats, as "half-endian" formats are also used. There's also the more philosophical point that endianness only comes into play at these input/output situations (and maybe for big int calculations) and it seems to me overkill to have as first-class types, particularly for floats (which are hard enough to get right anyway), and the simplification benefits (of removing them for floats at least) are significant.

[Kelimion]

Implementing file formats as I do, I have found the first class integer endian types to be very welcome. They make the structs and enums more self-documenting, instead of guessing which of them have to be byte-swapped when, and on which platforms.

As for endian-specific floats, this is something for @gingerBill to look into, but I'd not write them off just yet.

[gitlost]

Ha won't do!