On Fri, 28 Aug 2026 at 13:10, Steve Thompson <[email protected]> wrote:
> I am considering how to get software products to run using fewer MSUs.... > Products that you are writing, or someone else's existing code? > Is there such a thing as a functional Characteristics manual > these days that describes the behaviors of an instruction on/in a > particular machine level. > There's nothing like the old FC manuals that describe model instruction performance in detail, and hasn't been for decades. The Principles of Operation does have some performance suggestions scattered through it, but they're not connected to a particular model. The place I'd start is a series of documents, the first of which is "How Do You Do What You Do When You're a CPU?" by Bob Rogers, from about 2005. There was a series of updates to this as each new generation of machines came out, e.g. "How Do You Do What You Do When You're a z13 CPU?", and so on. Some of these papers were presented at SHARE sessions, and are available from the share.org site. I think some others were published independently by IBM and by Bob Rogers after he retired from IBM. The latest along this line that I'm aware of are "IBM Z / LinuxONE System Processor Optimization Primer" by Kevin Shum, and a follow-on "The IBM Z Microprocessor Optimization Primer" by David Hutton. That last is from 2023, so there might well be an update by now. Some Googling will yield most of them, and some related papers by other people, e.g. on the increasing perils of storing into the instruction stream (SIIS). While most of what they contain is not highly specific to particular models the way the FC manuals were, what you do see is the general movement away from classic S/370 implementations to the way modern CPUs actually work. Writing your code to generally honour those principles rather than trying to target the latest model is almost certainly the best approach. Leave the specific targeting to compiler writers. Example: At one time, I knew that using up to 6 MVC instructions > was actually faster than an MVCL. (Assuming I have remembered the > number of MVCs correctly). > This kind of thing changes from model to model. I wouldn't try to keep up - just follow the general principles. It may well be that today's CPU will turn an MVCL into something that performs better than the string of MVCs you wrote yesterday. And certainly a current compiler is likely to go well beyond that. Vector instructions are something a compiler might use that you probably haven't considered for string manipulation. Since IBM doesn't make PLAS (or PLX, or whatever it is called > today) available to us software developers I was just doing some > blue sky thinking and trying to decide if we have this kind of > info, if the juice would be worth the squeeze. > It's very unlikely to be worth it. If you have to write some intensely used small assembler routine, write it in C and see what the compiler generates with the latest TUNE and TARGET options. Then extract and adapt that if it has to be embedded in your high performance routine. Tony H. ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
