Reenable the use of the force-soft-floats feature

[antoyo]

This fixes an infinite recursion in sqrt for codegen backends that implement intrinsics like simd_fsqrt by calling sqrt on every element of the vector.

This is going to be needed to fix rust-lang/compiler-builtins#649.

I was able to test this locally by changing some other code (basically having an implementation of _mm_sqrt_pd that calls the local sqrt function on all elements of the vector).
Do you have any ideas for how to add a test for this? Perhaps by including this implementation of _mm_sqrt_pd behind a #[cfg]?

[tgross35]

I think this PR shouldn't be needed; usage of force-soft-floats in source was replaced with arch_enabled to simplify the logic, which select_implementation! already guards against

libm/src/math/support/macros.rs

Lines 74 to 75 in c9672e5

	// By default, never use arch-specific implementations if we have force-soft-floats
	#[cfg(arch_enabled)]

. The config is set here

libm/configure.rs

Lines 79 to 88 in c9672e5

	/// Simplify the feature logic for enabling arch-specific features so code only needs to use
	/// `cfg(arch_enabled)`.
	fn emit_arch_cfg() {
	println!("cargo:rustc-check-cfg=cfg(arch_enabled)");
	// Enabled by default via the "arch" feature, `force-soft-floats` overrides to disable.
	if cfg!(feature = "arch") && !cfg!(feature = "force-soft-floats") {
	println!("cargo:rustc-cfg=arch_enabled");
	}
	}

, compiler-builtins probably just needs to be fixed to have the same logic.

[antoyo]

I think this PR shouldn't be needed; usage of force-soft-floats in source was replaced with arch_enabled to simplify the logic, which select_implementation! already guards against

I thought arch_enabled would enable the usage of assembly. So won't this disable assembly in addition to hardware floats?

[tgross35]

(replying at rust-lang/compiler-builtins#810 since I had a better answer there)

[antoyo]

Closing since this was fixed elsewhere.