compiler: avoid negative zero in float constants