From f83f5ca2582c46d2df4d080dcb3719af0a517e58 Mon Sep 17 00:00:00 2001 From: lkcl Date: Fri, 5 May 2023 01:49:13 +0100 Subject: [PATCH] --- openpower/sv/rfc/ls012.mdwn | 22 +++++++++++++++++++++- 1 file changed, 21 insertions(+), 1 deletion(-) diff --git a/openpower/sv/rfc/ls012.mdwn b/openpower/sv/rfc/ls012.mdwn index 6b1ca3fde..9cf93f38d 100644 --- a/openpower/sv/rfc/ls012.mdwn +++ b/openpower/sv/rfc/ls012.mdwn @@ -646,7 +646,27 @@ instruction do too much. Another good example is the Integer Twin-butterfly instructions, `((a +/- b) * c) >> sh` which require **eight** instructions and temporary registers. Although expensive they save so many other -instructions it is hard to disregard them. +instructions - and registers - that it is hard to disregard them +even if their internal implementation is Micro-coded. + +**Is it general-purpose or have a compelling use-case?** + +The more specialised the instruction the less power used but the less +opportunity it has for being used elsewhere. Good examples of bad +instructions are illustrated by an MSc proposing a chacha20 SIMD add-xor-rotate-by-7 +instruction, when chacha20 has nowhere near the decades-established use as Rijndael +(AES) or SHA. Although the instruction halved the number of inline-unrolled +instructions in chacha20 it is clearly so specific as to be useless for any other purpose. + +Good examples of good specialist instructions are the +AES and SHA round-acceleration instructions in VSX, because these algorithms +are so heavily used that nearly all ISAs have them. + +Perhaps this point should be placed first but it is a different angle on +the cost-benefit analysis that starts with "Does anyone want it": that +alone is not quite enough, because although a given Stakeholder might want +a particular instruction to accelerate *their* application, the expression +of need is only where the evaluation process *begins*. **Summary** -- 2.30.2