[PATCH] D136464: [AArch64][SVE2] Add the SVE2.1 signed and unsigned 2-way dot instructions

David Sherwood via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Tue Oct 25 03:27:21 PDT 2022


david-arm marked an inline comment as done.
david-arm added inline comments.


================
Comment at: llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td:3585-3588
+def SDOT_2WAY_ZZZ_S  : sve2p1_two_way_dot_vv<"sdot", 0b0>;
+def UDOT_2WAY_ZZZ_S  : sve2p1_two_way_dot_vv<"udot", 0b1>;
+def SDOT_2WAY_ZZZI_S : sve2p1_two_way_dot_vvi<"sdot", 0b0>;
+def UDOT_2WAY_ZZZI_S : sve2p1_two_way_dot_vvi<"udot", 0b1>;
----------------
sdesmalen wrote:
> paulwalker-arm wrote:
> > paulwalker-arm wrote:
> > > sdesmalen wrote:
> > > > paulwalker-arm wrote:
> > > > > For clarity, can you create a follow up patch to ensure all the SVE dot instructions use this format? This presumably means adding `4WAY`.
> > > > We should be using VG2 or VG4 for that purpose, that aligns with what we've done for SME2
> > > VG2/4 relates to instruction operands that are multi-register.  Here 2WAY relates to the instruction's operation, which in this case only take single register operands.  For earlier versions of SVE there was no distinction because all the dot instructions were 4WAY.  This now changes with SVE2p1, hence my suggestion.
> > Just noting D135676 shows a different naming scheme (e.g. BtoH and BtoS).  This also works to distinguish between 2-way and 4-way so either are good options but we need to choose one and stick with it.
> > Here 2WAY relates to the instruction's operation
> Of course, I should have looked more closely before I posted.
> 
> > This also works to distinguish between 2-way and 4-way so either are good options but we need to choose one and stick with it.
> In that case, I'd prefer to align with D135676 and use `_HtoS`.
Renamed to be SDOT_ZZZ_HtoS, etc.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D136464/new/

https://reviews.llvm.org/D136464



More information about the llvm-commits mailing list