I know, it was an e-mail, not a white paper.

One type of optimization is simply arranging the instructions better. Along 
with the ARCH parameter there is a TUNE parameter.

> programs that may never get compiled again

And for those, IBM has the Automatic Binary Optimizer. 
https://www.ibm.com/us-en/marketplace/improved-cobol-performance 

Charles


-----Original Message-----
From: IBM Mainframe Assembler List [mailto:[email protected]] On 
Behalf Of Rob van der Heij
Sent: Wednesday, January 24, 2018 12:24 PM
To: [email protected]
Subject: Re: Fair comparison C vs HLASM

On 24 January 2018 at 18:33, Charles Mills <[email protected]> wrote:

>
> The reality is that cycle times are not getting any faster. A z14 does 
> not execute z10 machine instructions significantly (any?) faster than a z10.
>

The second sentence does not follow from the first one. While a single 
instruction may take (in many cases) the same number of clock cycles, improved 
pipeline and out-of-order execution with deeper and wider cache makes that a 
series of instructions often does run quicker than before.
Workloads that just didn't fit in cache anymore before may well fit now.
I'm frequently surprised by the number of extra instructions that you can sneak 
into the code without slowing it down. Add to that the SMT support for 
specialty engines.

One of the reasons for investing in those aspects of a CPU is that a fair 
amount of the instructions executedare in programs that may never get compiled 
again. Would be interesting to know which percentage of the executed 
instructions are plain old S/370 ones...

Rob

Reply via email to