The hypotf implementation for PowerPC (sysdeps/powerpc/fpu/e_hypotf.c) generates spurious overflow exception for subnormal arguments. This is due it checks if arg1/arg2 > 2^30 (arg1 > arg2) by using a FP division. Since the double version of ieee sqrt function will be used to compute final value, there is not need to argument reduction in such case. A possible fix would be just to clean up the test and call ieee sqrt directly. I will send a patch to libc-alpha with a fix.
Fixed by commit 16e616a72f9ac247520c0c7da99b99e229facdf9.