[PATCH] D35566: [DAG] Extend visitSCALAR_TO_VECTOR optimization to truncated vector.

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: lib/CodeGen/SelectionDAG/DAGCombiner.cpp:15628
     ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
-    if (C0 && VT == InVec.getValueType() &&
-        VT.getScalarType() == InVal.getValueType()) {
-      SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
+    if (C0 && VT.getScalarType() == InVal.getValueType()) {
+      SmallVector<int, 8> NewMask(InVecT.getVectorNumElements(), -1);
----------------
The only way for `VT.getScalarType() == InVal.getValueType()` to be false is if legalization promoted it to a larger integer type... and I can't why that should block the transform (the extra bits get discarded anyway).


================
Comment at: lib/CodeGen/SelectionDAG/DAGCombiner.cpp:15635
                                     NewMask);
+      if (VT.getVectorNumElements() <= InVecT.getVectorNumElements() &&
+          TLI.isShuffleMaskLegal(NewMask, VT)) {
----------------
Do you need to check that the element type matches?  And in that case, isn't the "VT.getVectorNumElements() == InVecT.getVectorNumElements()" check redundant (since it implies `VT == InVecT`)?


https://reviews.llvm.org/D35566