Usage of core library operations in generic implementations

[BilelGho]

I am not sure if this goes against the expected behavior from generic implementations. Please feel free to dismiss this if it is the case:

By a look at the generic implementation for sqrt

compiler-builtins/libm/src/math/generic/sqrt.rs

Lines 110 to 111 in 599f0e6

	// Normalize subnormals by multiplying by 1.0 << SIG_BITS (e.g. 0x1p52 for doubles).
	let scaled = x * F::from_parts(false, F::SIG_BITS + F::EXP_BIAS, zero);

It can be seen that the current implementation relies on the core implementation for floating point multiplication, which itself may compile to a hardware implementation, even if libm is being compiled with force-soft-floats enabled.

[tgross35]

Thanks for bringingthis up, but it is expected :) force-soft-floats is a bit of a bad name inherited from some old code. All it does is prevent using assembly or intrinsics because some backends historically didn't support that well, but it can use simple ops fine.

I'd actually like to remove this feature, but that's blocked on a few things.