Add builtins for f16/f128 float conversions

[beetrees]

Adds builtins for float conversions to/from f16/f128, which will solve rust-lang/rust#123885. I also updated the README.md to reflect that f16/f128 are being added to Rust.

@tgross35 this should save you some work on #587.

I'm not sure if the new functions need the #[avr_skip] attribute like the existing extend/truncate builtins do, but thought it was better to add it and accidentally cause linker errors if the builtin is needed and not supplied by libgcc rather than not having the attribute and accidentally causing subtle ABI bugs if a special ABI is required on AVR. @Patryk27 who added the original #[avr_skip]s.

[Patryk27]

Yeah, I'm not sure on AVR's ABI here, so probably better to keep those #[avr_skip] 🙂

[tgross35]

Should this also be setting COMPILER_RT_HAS_FLOAT16 for the C build, in case that winds up being used as a fallback in the future?

[tgross35]

Since this will probably land before #587, you might have to add the feature gate that @bjorn3 requested #587 (comment)

[beetrees]

I've added a no-f16-f128 feature (following the convention set by the existing no-asm feature). I've also added a COMPILER_RT_HAS_FLOAT16 define when compiler-rt is being built, with a comment explaining why it is there.

[tgross35]

Thanks, that part looks great to me. Looks like tests might be missing? Like

compiler-builtins/testcrate/tests/misc.rs

Lines 114 to 121 in 2978bdd

	// PowerPC tests are failing on LLVM 13: https://github.com/rust-lang/rust/issues/88520
	#[cfg(not(target_arch = "powerpc64"))]
	#[test]
	fn float_extend() {
	use compiler_builtins::float::extend::__extendsfdf2;
	extend!(f32, f64, __extendsfdf2);
	}

/

compiler-builtins/testcrate/tests/misc.rs

Lines 201 to 214 in 2978bdd

	#[test]
	fn float_trunc() {
	use compiler_builtins::float::trunc::__truncdfsf2;
	trunc!(f64, f32, __truncdfsf2);
	}
	#[cfg(target_arch = "arm")]
	#[test]
	fn float_trunc_arm() {
	use compiler_builtins::float::trunc::__truncdfsf2vfp;
	trunc!(f64, f32, __truncdfsf2vfp);
	}

[beetrees]

As the builtins aren't present everywhere at the moment (and therefore comparing against the builtin compiler float conversions would cause test failures on platforms where they're currently unavailable or miscompiled), I've instead added tests comparing compiler_builtins float conversions to those in rustc_apfloat, the (heavily tested) soft float library ported from LLVM for use in rustc. I've also fixed a copy-paste error and some shift overflow bugs (that only show up now the destination type can be larger than 32 bits) in trunc.rs that I discovered by running the tests, and moved the float conversion tests from misc.rs to conv.rs.

[tgross35]

I've instead added tests comparing compiler_builtins float conversions to those in rustc_apfloat

Great idea, I can probably do the same on #587. Everything here looks reasonable to my understanding, probably ready for a look from @Amanieu.

[Amanieu]

I'm having trouble understand how exactly this helps prevent overflows. Aren't you casting to a smaller size (u16) here instead of operating as a u32?

[beetrees]

@Amanieu This is indeed not an issue for truncation to f16. However, the __trunctfdf2 builtin also implemented in this PR truncates from f128 to f64, meaning R::Int is a u64 and R::SIGNIFICAND_BITS is 52.

[Amanieu]

Ah, I missed that. LGTM.