<<Even those (old) PowerPC architectures where all computations are done in double precision should have single-precision IEEE arithmetic fully supported in GCC; GCC will add appropriate rounding instructions when necessary. >> That's not quite good enough, due to double rounding phenomena