[llvm] [SelectionDAG][x86] Ensure vector reduction optimization (PR #144231)

[Simon Pilgrim via llvm-commits]

Suhajda =?utf-8?q?Tamás?= <sutajo at gmail.com>
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/144231 at github.com>


================
@@ -48066,6 +48077,36 @@ static SDValue combineExtractVectorElt(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static SDValue combineExtractVectorEltAndOperand(SDNode* N, SelectionDAG& DAG,
+    TargetLowering::DAGCombinerInfo& DCI,
+    const X86Subtarget& Subtarget)
+{
+  bool TransformedBinOpReduction = false;
+  auto Op = combineExtractVectorElt(N, DAG, DCI, Subtarget, TransformedBinOpReduction);
+
+  if (TransformedBinOpReduction)
+  {
+    // In case we simplified N = extract_vector_element(V, 0) with Op and V
+    // resulted from a reduction, then we need to replace all uses of V with
+    // scalar_to_vector(Op) to make sure that we eliminated the binop + shuffle
+    // pyramid. This is safe to do, because the elements of V are undefined except 
+    // for the zeroth element and Op does not depend on V.
+
+    auto OldV = N->getOperand(0);
----------------
RKSimon wrote:

(style) don't use auto

https://github.com/llvm/llvm-project/pull/144231