[X86] Legalize (v64i1 (bitcast (i64 X))) on 32-bit targets by extracting 32-bit halves from i32, bitcasting each to v32i1, and concatenating.

This prevents the scalarization that would otherwise occur.

llvm-svn: 324057

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -23602,6 +23602,22 @@ static SDValue LowerBITCAST(SDValue Op, const X86Subtarget &Subtarget,
   MVT SrcVT = Op.getOperand(0).getSimpleValueType();
   MVT DstVT = Op.getSimpleValueType();
 
+  // Legalize (v64i1 (bitcast i64 (X))) by splitting the i64, bitcasting each
+  // half to v32i1 and concatenating the result.
+  if (SrcVT == MVT::i64 && DstVT == MVT::v64i1) {
+    assert(!Subtarget.is64Bit() && "Expected 32-bit mode");
+    assert(Subtarget.hasBWI() && "Expected BWI target");
+    SDValue Op0 = Op->getOperand(0);
+    SDLoc dl(Op);
+    SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0,
+                             DAG.getIntPtrConstant(0, dl));
+    Lo = DAG.getBitcast(MVT::v32i1, Lo);
+    SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0,
+                             DAG.getIntPtrConstant(1, dl));
+    Hi = DAG.getBitcast(MVT::v32i1, Hi);
+    return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi);
+  }
+
   if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8 ||
       SrcVT == MVT::i64) {
     assert(Subtarget.hasSSE2() && "Requires at least SSE2!");