Before I joined IBM I worked at the Stanford Linear Accelerator
Center, where most of the production code was written in Fortran.
I once asked an IBM rep what benchmarks they used for testing
new architecture against Fortran code, and was told that the code
had been generated by Fortran G (non-optimizing) -- this, years
after the optimizing Fortran H compiler was widely used.
Those list contributors with performance-sensitive applications
might consider donating "kernels" to IBM as examples of real-world
workloads.
John Ehrman
(------------------ Referenced Note Follows --------------------)
Date: 8 October 2010, 08:09:55 -0700
From: IBM Mainframe Assembler List ASSEMBLE at INTERNET
On 10/8/2010 7:51 AM, Blaicher, Chris wrote:
> Yes, I use MVCL, TRT, CLCL and the rest of the millicoded instructions,
I just use them where appropriate.
The irony here is that IBM uses sampling of real workloads to
determine which instructions are used most. That knowledge helps them
decide which instructions should be given preferential performance
treatment and which should not. So, if an instruction isn't often
used because it's slow, it will never get any faster...
