[PATCH] D111657: [SVE][CodeGen] Enable reciprocal estimates for scalable fdiv/fsqrt

[Paul Walker via Phabricator via llvm-commits]

paulwalker-arm added a comment.

Just a couple of extra points that depend on the answer to my previous question.



================
Comment at: llvm/lib/Target/AArch64/SVEInstrFormats.td:1935-1936
 
+  def : SVE_2_Op_Pat<nxv2f16, op, nxv2f16, nxv2f16, !cast<Instruction>(NAME # _H)>;
+  def : SVE_2_Op_Pat<nxv4f16, op, nxv4f16, nxv4f16, !cast<Instruction>(NAME # _H)>;
   def : SVE_2_Op_Pat<nxv8f16, op, nxv8f16, nxv8f16, !cast<Instruction>(NAME # _H)>;
----------------
If the answer to my previous question is that it is unsafe for unpacked vectors then please remove these patterns.  If there comes a time that we definitely want to support unpacked vectors I think we'll probably need slightly differ ISEL patterns for those types.


================
Comment at: llvm/lib/Target/AArch64/SVEInstrFormats.td:2637
 
+  def : SVE_1_Op_Pat<nxv2f16, op, nxv2f16, !cast<Instruction>(NAME # _H)>;
+  def : SVE_1_Op_Pat<nxv4f16, op, nxv4f16, !cast<Instruction>(NAME # _H)>;
----------------
As above.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D111657/new/

https://reviews.llvm.org/D111657