[PATCH] D114951: [Analysis][AArch64] Add on the address computational cost for gathers/scatters

[David Sherwood via Phabricator via llvm-commits]

david-arm abandoned this revision.
david-arm added a comment.

I'm abandoning this patch for now as it requires more investigation based on the comments. There are inconsistencies in the vectoriser, for example when vectorising a normal load (VF>1) we don't add on the address computation cost, whereas do add on the cost for scalar loads (VF=1). This clearly leads to the vectoriser being biased towards vectorisation for AArch64 targets! I think it's worth probably first making the vectoriser consistent in this behaviour, i.e. the we have costs like this:

LV: Found an estimated cost of 2 for VF 1 For instruction:   %1 = load double, double* %arrayidx.us, align 8, !tbaa !8
LV: Found an estimated cost of 1 for VF 2 For instruction:   %1 = load double, double* %arrayidx.us, align 8, !tbaa !8


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D114951/new/

https://reviews.llvm.org/D114951