LOCTR is your friend.
-- Shmuel (Seymour J.) Metz http://mason.gmu.edu/~smetz3 עַם יִשְׂרָאֵל חַי נֵ֣צַח יִשְׂרָאֵ֔ל לֹ֥א יְשַׁקֵּ֖ר ________________________________________ From: IBM Mainframe Discussion List <[email protected]> on behalf of Bernd Oppolzer <[email protected]> Sent: Saturday, August 29, 2026 4:42 AM To: [email protected] <[email protected]> Subject: Re: Efficiency of Instructions External Message: Use Caution "Splitting data from Instruction areas (I-Bank/D-Bank) is really needed ..." this is known as SIIS ... store into instruction stream. I had to work on some my customer's very old Assembler routines (from the early 80's) in 2021. The routines are at the base of the IMS dialog framework and performance degraded heavily, when they switched to a new machine model. The Assembler routines showed heavy cache losses, and we had to fix them due to these effects. There were 4 major issues: 1. home written structered programming macros (IF, THEN, ELSE) ... especially LOOP, which stored its loop control variable near the instruction stream :-) 2. other home grown macros which stored their variables near the I-stream (using the *+offset approach) 3. or IBM macros, doing the same 4. home grown "procedure" macros, which have their local (static) variables just behind the instructions; so the distance to the I-stream is too small What I did (in short): - creating a DSECT in the overall startup macro, where the other macros can put their variables into, and allocating the space for it later, far away from the I-stream (this resolves 1 and 2) - create home grown solutions for the IBM macros, TIME comes to mind - this resolves 3 - create some space (512 bytes) after the end of the procedure and before the variables start (subtract the length of the created literals before) - resolves 4 In the end, we resolved the cache losses and performance returned to "normal". HTH, kind regards Bernd Am 28.08.2026 um 21:32 schrieb Steve Thompson: > Well, given the amount of I/O the product does, it is sensitive to > various changes that IBM has done. Splitting data from Instruction > areas (I-Bank/D-Bank) is really needed from my experience working on > old ALC code written from Base S/370 through ESA. Amazing what speed > increases you get when you stop having contaminated cache lines. Think > in terms of OPEN in line with other code that is NOT using a work > area. The same for PUT/GET, etc. > > ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
