What you are saying is only partly true. Yes, the packed decimal instructions are horrible today from performance perspective, because they require direct access to memory.

However, recently, IBM has added decimal floating-point arithmetic to z/Architecture CPUs (and also made it part of updated IEEE 754 standard). These modern instructions do not suffer from the problems of old instructions, and in fact should be preferred to packed decimal for applications that need decimal point arithmetic.

And more broadly, IBM still adds a lot of specialized instructions to mainframe CPUs. For example, calls to cryptographic accelerators.