Change hardcoded `_FloatN` support to detect instead

[tgross35]

It looks like the f16 tests are disabled because of the CI GCC version

abi-cafe/include/harness/abi-cafe-rules.toml

Lines 17 to 19 in 94d3803

	# CI GCC is too old to support _Float16
	[target.x86_64-unknown-linux-gnu."f16::conv_c"]
	random = true

, and support is hardcoded at

abi-cafe/src/toolchains/c/declare.rs

Lines 117 to 160 in 94d3803

	match &self.cc_flavor {
	CCFlavor::Gcc
	if cfg!(any(
	target_arch = "x86",
	target_arch = "x86_64",
	target_arch = "aarch64",
	target_arch = "riscv32",
	target_arch = "riscv64",
	target_arch = "loongarch64",
	// GCC PowerPC support requires the VSX feature, which is only
	// enabled by default on powerpc64le. Rust doesn't currently support
	// `cfg(target_feature = "vsx").
	all(target_arch = "powerpc64", target_endian = "little"),
	)) && !cfg!(target_vendor = "apple") =>
	{
	"_Float128 "
	}
	CCFlavor::Gcc => Err(UnsupportedError::Other(
	"GCC isn't known to support f128 on this target".to_owned(),
	))?,
	CCFlavor::Clang | CCFlavor::Zigcc
	if cfg!(any(
	target_arch = "x86",
	target_arch = "x86_64",
	target_arch = "aarch64",
	target_arch = "riscv32",
	target_arch = "riscv64",
	// Clang PowerPC support requires the VSX feature, which is only
	// enabled by default on powerpc64le. Rust doesn't currently support
	// `cfg(target_feature = "vsx").
	all(target_arch = "powerpc64", target_endian = "little"),
	)) && !cfg!(any(
	target_env = "msvc",
	target_vendor = "apple",
	)) =>
	{
	"__float128 "
	}
	CCFlavor::Clang | CCFlavor::Zigcc => Err(UnsupportedError::Other(
	"Clang isn't known to support f128 on this target".to_owned(),
	))?,
	CCFlavor::Msvc => Err(UnsupportedError::Other(
	"MSVC doesn't support f128".to_owned(),
	))?,

. Could this change to be autodetected instead?

The most obvious way to do this would probably be to test for success/failure of echo "_Float16 x;" | gcc -x c -c - -o/dev/null. Alternatively, run: GCC/Clang with -E -dM -x c /dev/null -o - to grab preprocessor output for this source:

#define __STDC_WANT_IEC_60559_TYPES_EXT__
#include <float.h>

And look for the following defines:

• FLT16_MANT_DIG for _Float16
• FLT128_MANT_DIG for _Float128
• __SIZEOF_FLOAT128__ for __float128, or on Clang __FLOAT128__

FLT16_IS_IEC_60559 = 1 / FLT128_IS_IEC_60559 = 1 would probably be better than MANT_DIG since they allow you to verify the format is IEEE754, but they're not yet set by Clang.