[llvm] [AMDGPU] Legalize 64bit elements for BUILD_VECTOR on gfx942 (PR #145052)

[Matt Arsenault via llvm-commits]

================
@@ -17497,6 +17536,95 @@ SDValue SITargetLowering::performSelectCombine(SDNode *N,
                          SelectLHS, SelectRHS);
 }
 
+SDValue
+SITargetLowering::performBuildVectorCombine(SDNode *N,
+                                            DAGCombinerInfo &DCI) const {
+  // TODO: legalize for all targets instead of just v_mov_b64 enabled ones,
+  // legalizing could still enable s_mov_b64 which is supported on all targets.
+  const GCNSubtarget *ST = getSubtarget();
+  if (DCI.Level < AfterLegalizeDAG || !ST->hasMovB64())
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc SL(N);
+
+  EVT VT = N->getValueType(0);
+  EVT EltVT = VT.getVectorElementType();
+  unsigned SizeBits = VT.getSizeInBits();
+  unsigned EltSize = EltVT.getSizeInBits();
+
+  // Skip if:
+  //  - Value type isn't multiple of 64 bit (e.g., v3i32), or
+  //  - Element type has already been combined into 64b elements
+  if ((SizeBits % 64) != 0 || EltVT == MVT::i64 || EltVT == MVT::f64)
----------------
arsenm wrote:

Go of scalar bit size == 64

https://github.com/llvm/llvm-project/pull/145052