From: lkcl <lkcl@web>
Date: Fri, 23 Sep 2022 13:53:03 +0000 (+0100)
Subject: (no commit message)
X-Git-Tag: opf_rfc_ls005_v1~317
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=26c176ef83ef47caf8dea9442318ce1fce8a6eb0;p=libreriscv.git

---

diff --git a/openpower/sv/overview/discussion.mdwn b/openpower/sv/overview/discussion.mdwn
index 891dd6ee7..8eebff62c 100644
--- a/openpower/sv/overview/discussion.mdwn
+++ b/openpower/sv/overview/discussion.mdwn
@@ -102,3 +102,33 @@ The "correct" place to add such reversing would be the `REMAP`
 subsystem but the number of available bits is already under
 pressure so some care is needed. Even better would be the 24-bit
 RM but that is also precious and under pressure.
+
+*In theory* there is one bit free:
+
+| 0-1 |  2  |  3   4  |  description              |
+| --- | --- |---------|-------------------------- |
+| 00  |   0 |  dz  sz | simple mode                      |
+| 00  |   1 | 0  RG   | scalar reduce mode (mapreduce) |
+| 00  |   1 | 1  /    | reserved     |
+
+which could be utilised to instruct hardware between the
+regfile and the Arithmetic ALUs to perform **element*-level
+byte-reversing:
+
+    GPR -> brev(RM.normal.br) -> add -> brev(RM.normal.br) -> GPR
+
+However even this does not solve the problem caused by loading
+the data in 8-byte (`ld/ldbrx`) followed by **accessing** it
+as element-width-overridden half/word/byte elements: the situation
+occurs in the 8/16 table, above.
+
+In addition, it would **still be necessary** to instruct Hardware
+Architects on ensuring that Memory-Load to Regfile-byte-order is
+still **strictly** defined (architecturally, not actual implementation)
+
+Then also there is the completely separate issue of how to describe this
+in MSB0 numbering, which becomes a nightmare all on its own: one that
+has to be solved in the ISACaller Simulator when elwidth overrides are
+completed (recall that the ISACaller Simulator uses a python class where
+numbers are indeed strict MSB0 defined, arithmetically).
+