It really comes down to which architectures you're porting to. The two biggest issues are big endian vs little endian, and memory consistency models. Little endian is the clear winner for actively-developed architectures, but there are still plenty of vintage big endian architectures to target, and it looks like IBM mainframes at least are still exclusively big endian.
For memory consistency, Alpha historically had value as the weakest and most likely to expose bugs. But nobody really wants to implement hardware like that anymore, almost everything falls somewhere on the spectrum of behavior bounded by x86 (strict) and Arm (weaker), and newer languages (eg. C++ 11) mean newer code can be explicit about its expectations rather than ambiguous or implicit.
For memory consistency, Alpha historically had value as the weakest and most likely to expose bugs. But nobody really wants to implement hardware like that anymore, almost everything falls somewhere on the spectrum of behavior bounded by x86 (strict) and Arm (weaker), and newer languages (eg. C++ 11) mean newer code can be explicit about its expectations rather than ambiguous or implicit.