[llvm] [AMDGPU][SDAG] Legalise v2i32 or/xor/and instructions to make use of 64-bit wide instructions (PR #140694)

[Chris Jackson via llvm-commits]

================
@@ -4056,6 +4056,53 @@ SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
   SDLoc SL(N);
   SelectionDAG &DAG = DCI.DAG;
 
+  // When the shl64_reduce optimisation code is passed through vector
+  // legalization some scalarising occurs. After ISD::AND was legalised, this
+  // resulted in the AND instructions no longer being elided, as mentioned
+  // below. The following code should make sure this takes place.
+  if (RHS->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
+    SDValue VAND = RHS.getOperand(0);
+    if (ConstantSDNode *CRRHS = dyn_cast<ConstantSDNode>(RHS->getOperand(1))) {
+      uint64_t AndIndex = RHS->getConstantOperandVal(1);
+      if (VAND->getOpcode() == ISD::AND && CRRHS) {
+        SDValue LHSAND = VAND.getOperand(0);
+        SDValue RHSAND = VAND.getOperand(1);
+        if (RHSAND->getOpcode() == ISD::BUILD_VECTOR) {
+          // Part of shlcombine is to optimise for the case where its possible
+          // to reduce shl64 to shl32 if shift range is [63-32]. This
+          // transforms: DST = shl i64 X, Y to [0, shl i32 X, (Y & 31) ]. The
+          // '&' is then elided by ISel. The vector code for this was being
+          // completely scalarised by the vector legalizer, but now v2i32 is
+          // made legal the vector legaliser only partially scalarises the
+          // vector operations and the and was not elided. This check enables us
+          // to locate and scalarise the v2i32 and and re-enable ISel to elide
+          // the and instruction.
+          ConstantSDNode *CANDL =
+              dyn_cast<ConstantSDNode>(RHSAND->getOperand(0));
+          ConstantSDNode *CANDR =
+              dyn_cast<ConstantSDNode>(RHSAND->getOperand(1));
+          if (CANDL && CANDR && RHSAND->getConstantOperandVal(0) == 0x1f &&
+              RHSAND->getConstantOperandVal(1) == 0x1f) {
+            // Get the non-const AND operands and produce scalar AND
+            const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
+            const SDValue One = DAG.getConstant(1, SL, MVT::i32);
+            SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32,
+                                     LHSAND, Zero);
+            SDValue Hi =
+                DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, LHSAND, One);
+            SDValue AndMask = DAG.getConstant(0x1f, SL, MVT::i32);
----------------
chrisjbris wrote:

I think I understand now. I've removed the duplication of the constant node.

https://github.com/llvm/llvm-project/pull/140694