[Libre-soc-isa] [Bug 570] svp64 vector loads: sub-dword selection before or after byte-reversal

bugzilla-daemon at libre-soc.org bugzilla-daemon at libre-soc.org
Thu Jan 7 00:17:10 GMT 2021


https://bugs.libre-soc.org/show_bug.cgi?id=570

--- Comment #1 from Luke Kenneth Casson Leighton <lkcl at lkcl.net> ---
(In reply to Alexandre Oliva from comment #0)
> Last night, while going over
> https://libre-soc.org/simple_v_extension/appendix/ with a particular focus
> on ld's operation with an elwidth overrider for the src, I missed various
> details in the specification.

apologies i should have mentioned, that was the older version of the spec,
relevant exclusively to RISC-V. given that RV had BE removed at around version
3 (RISC-III) it was not discussed, at all.

the reason i referred to that older spec was to illustrate to Cole that there
did exist walkthroughs for twin element width overrides.



> whole register.  this suggested to me that the sub-indexing of the value
> loaded memory would take place after byte-reversion.

the pseudocode as modified and derived in the other bugreport will be the
correct pseudocode.

what you are looking at in the RV variant is unfortunately not relevant, i
apologise, as far as bytereversal is concerned, only elwidths and extension.

given that this is the case can i recommend closing this one and starting
again, from the pseudocode listed in 567

https://bugs.libre-soc.org/show_bug.cgi?id=567#c2

unfortunately almost all of what you wrote is invalid when viewed without the
addition of the (fully OpenPOWER v3.0B Compliant) ld/brx-LE/BE bytereversing
that gets everything into NEON-style internal representation and ordering as
far as Vectorisation is concerned.

i will go over it thoroughly and make sure nothing was missed but,
realistically, we need to close this one as invalid.  sorry about that.

-- 
You are receiving this mail because:
You are on the CC list for the bug.


More information about the Libre-SOC-ISA mailing list