Microoptimization for code size in floating point comparisons on x86-64
This affects EQL on real and complex floats and "=" on floats when at
least one of the two arguments is complex.
Use a 32-bit register as the destination of the MOVMSKP[SD] and the
source of the integer comparison instead of a 64-bit one. This doesn't
change the semantics but makes both instruction's encodings shorter.
If the register is EAX do the comparison on AL as this additionally
shortens the encoding.