[PATCH] D140789: [SLP] Unify GEP cost modeling for load, store and GEP nodes.

[Alexey Bataev via Phabricator via llvm-commits]

ABataev added inline comments.


================
Comment at: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6901-6907
+      // For any external uses of some lanes extract element instructions will
+      // be generated (which cost is estimated separately). For the latter case
+      // since the set of GEPs itself is not vectorized those used more than
+      // once will remain staying in vectorized code as well. So we should not
+      // count them as savings.
+      if (!Ptr->hasOneUse() && isa<LoadInst, StoreInst>(VL0))
+        continue;
----------------
Probably not related to this patch. Do we need to emit extractelement for GEP at all? Is not it better just to rebuild the scalar GEP?


================
Comment at: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6906-6907
+      // count them as savings.
+      if (!Ptr->hasOneUse() && isa<LoadInst, StoreInst>(VL0))
+        continue;
+
----------------
Shall we drop this check? We still vectorize GEPs with multiple uses and then emit extractelement for them. The cost of the extractelement is calculated separately. So, when we calculate the cost for GEPs with multiple uses, we exclude them from saving cost and then we add an extra cost for extractelement. If we're still going to emit extractelement, need to remove this check (the original пуз will be vectorized and removed and then the extractelement is generated). If it is better to keep scalar copy, need to remove the cost of the extractelement calculation and keep this check.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D140789/new/

https://reviews.llvm.org/D140789