Guard truncate from vector float to vector __bf16 with !flag_rounding_math && HONOR_N...

Guard truncate from vector float to vector __bf16 with !flag_rounding_math && HONOR_NANS (BFmode).

hw instruction doesn't raise exceptions, turns sNAN into qNAN quietly,
and always round to nearest (even). Output denormals are always
flushed to zero and input denormals are always treated as zero. MXCSR
is not consulted nor updated.
W/o native instructions, flag_unsafe_math_optimizations is needed for
the permutation instructions.
Similar guard extend from vector __bf16 to vector float with
!HONOR_NANS (BFmode).

gcc/ChangeLog:

	* config/i386/i386.md (truncsf2bf2): Add !flag_rounding_math
	to the condition, require flag_unsafe_math_optimizations when
	native instruction is not available.
	* config/i386/mmx.md: (truncv2sfv2bf2): Ditto.
	(extendv2bfv2sf2): Add !HONOR_NANS (BFmode) to the condition.
	* config/i386/sse.md: (truncv4sfv4sf2): Add
	!flag_rounding_math to the condition, require
	flag_unsafe_math_optimizations when native instruction is not
	available.
	(truncv8sfv8bf2): Ditto.
	(truncv16sfv16bf2): Ditto.
	(extendv4bfv4sf2): Add !HONOR_NANS (BFmode) to the condition.
	(extendv8bfv8sf2): Ditto.
	(extendv16bfv16sf2): Ditto.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/avx512bf16-truncsfbf.c: Add -ffast-math.
	* gcc.target/i386/avx512bw-extendbf2sf.c: Ditto.
	* gcc.target/i386/avx512bw-truncsfbf.c: Ditto.
	* gcc.target/i386/sse2-extendbf2sf.c: Ditto.
	* gcc.target/i386/ssse3-truncsfbf.c: Ditto.