Consider using the `cpufeatures` crate

[str4d]

Every call to has_avx2 performs a runtime check:

rust-fn-dsa/fn-dsa-comm/src/lib.rs

Lines 229 to 260 in 8d01465

	/// Do a rutime check for AVX2 support (x86 and x86_64 only).
	///
	/// This is a specialized subcase of the is_x86_feature_detected macro,
	/// except that this function is compatible with `no_std` builds.
	#[cfg(all(not(feature = "no_avx2"),
	any(target_arch = "x86_64", target_arch = "x86")))]
	pub fn has_avx2() -> bool {
	#[cfg(target_arch = "x86_64")]
	use core::arch::x86_64::{__cpuid, __cpuid_count, _xgetbv};
	#[cfg(target_arch = "x86")]
	use core::arch::x86::{__cpuid, __cpuid_count, _xgetbv};
	unsafe {
	// Check that we can access function parameter 7 (where the AVX2
	// support bit resides).
	let r = __cpuid(0);
	if r.eax < 7 {
	return false;
	}
	// Check that AVX2 is supported by the CPU.
	let r = __cpuid_count(7, 0);
	if (r.ebx & (1 << 5)) == 0 {
	return false;
	}
	// Check that the full-size (256-bit) ymm registers are enabled.
	let r = _xgetbv(0);
	return (r & 0x06) == 0x06;
	}
	}

The cpufeatures crate provides an equivalent no_std-compatible alternative to is_x86_feature_detected!, that comes with built-in caching (in an AtomicU8), and elimination of the runtime checks if AVX2 is enabled at compile time:

cpufeatures::new!(cpuid_avx2, "avx2");

pub fn has_avx2() -> bool {
    cpuid_avx2::get()
}

(or just use the macro-created module / its InitToken directly).

[pornin]

Well I added that: dc0ee52
I am not entirely sure what it brings in practice in this case (apart from an extra dependency), since I was already using the test only in top-level functions (so the cost was negligible). But it seems harmless enough.

[pornin]

I guess I'll update that when (if) RFC 2725 is stabilized.