[PATCH] D8943: Calculate vectorization factor using the narrowest type instead of widest type

[hfinkel@anl.gov via llvm-commits]

hfinkel added inline comments.

================
Comment at: lib/Target/X86/X86TargetTransformInfo.cpp:903-905
@@ -902,4 +902,5 @@
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
+  auto VT = TLI->getValueType(DL, SrcVTy);
   int Cost = 0;
-  if (LT.second != TLI->getValueType(DL, SrcVTy).getSimpleVT() &&
+  if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
       LT.second.getVectorNumElements() == NumElem)
----------------
In that case, please do commit it.

================
Comment at: lib/Transforms/Vectorize/LoopVectorize.cpp:5166
@@ +5165,3 @@
+        auto TypeSize =
+            (unsigned)DL.getTypeSizeInBits(Inst->getType()->getScalarType());
+        RegUsage += std::max<unsigned>(1, VFs[j] * TypeSize / WidestRegister);
----------------
Does this do the right thing for stores? For stores, you want the type of the stored value, right?


http://reviews.llvm.org/D8943