Hi, I found a similar case (x close to 1 and large y) for double precision: x = 0x1.fffffd2e3e669p-1 y = 0x1.344c9823eb66cp+32 I find an error of 427.893 ulps with Newlib 3.3.0, whereas all other inputs I tried have an error of less than 1 ulp. Best regards, Paul