[llvm] [AMDGPU][X86][DAG] Avoid duplicate BinOp result from narrowing insert-extract sub-vector (PR #201056)

[Matt Arsenault via llvm-commits]

================
@@ -27267,36 +27295,106 @@ static SDValue narrowInsertExtractVectorBinOp(EVT SubVT, SDValue BinOp,
 
   EVT VecVT = BinOp.getValueType();
   SDValue Bop0 = BinOp.getOperand(0), Bop1 = BinOp.getOperand(1);
-  if (VecVT != Bop0.getValueType() || VecVT != Bop1.getValueType())
+  if (VecVT != Bop0.getValueType() || VecVT != Bop1.getValueType() ||
+      VecVT.isScalableVT())
     return SDValue();
-  if (!TLI.isOperationLegalOrCustom(BinOpcode, SubVT, LegalOperations))
+
+  // This fold only pays off when the wide binop disappears completely, so every
+  // user must be an extract_subvector. Require them all to extract N's type so
+  // a single chain scan serves every extract, and seed the source map by index.
+  EVT SubVT = N->getValueType(0);
+  if (VecVT.getSizeInBits() <= SubVT.getSizeInBits())
     return SDValue();
 
-  SDValue Sub0 = getSubVectorSrc(Bop0, Index, SubVT);
-  SDValue Sub1 = getSubVectorSrc(Bop1, Index, SubVT);
+  SmallMapVector<unsigned, std::tuple<SDNode *, SDValue, SDValue>, 4> Extracts;
----------------
arsenm wrote:

Feels like this shouldn't be a MapVector. You only have a small number of indexes to fill?

https://github.com/llvm/llvm-project/pull/201056