[AArch64][GlobalISel] Create copy rather than single-element concat

The verifier does not accept single-element G_CONCAT_VECTORS, so if there is a
single Op generate a COPY instead.

Author: davemgreen

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

@@ -446,7 +446,10 @@ void CombinerHelper::applyCombineShuffleConcat(MachineInstr &MI,
     }
   }
 
-  Builder.buildConcatVectors(MI.getOperand(0).getReg(), Ops);
+  if (Ops.size() > 1)
+    Builder.buildConcatVectors(MI.getOperand(0).getReg(), Ops);
+  else
+    Builder.buildCopy(MI.getOperand(0).getReg(), Ops[0]);
   MI.eraseFromParent();
 }
 

llvm/test/CodeGen/AArch64/GlobalISel/combine-shufflevector.mir

@@ -200,3 +200,31 @@ body:             |
     $q0 = COPY %z(<16 x s8>)
     RET_ReallyLR implicit $q0
 ...
+
+---
+name:            single_vector_to_copy
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $q0, $q1
+
+    ; CHECK-LABEL: name: single_vector_to_copy
+    ; CHECK: liveins: $q0, $q1
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %p1:_(<4 x s32>) = COPY $q0
+    ; CHECK-NEXT: %p2:_(<4 x s32>) = COPY $q1
+    ; CHECK-NEXT: $q0 = COPY %p1(<4 x s32>)
+    ; CHECK-NEXT: $q1 = COPY %p2(<4 x s32>)
+    ; CHECK-NEXT: RET_ReallyLR implicit $q0
+    %p1:_(<4 x s32>) = COPY $q0
+    %p2:_(<4 x s32>) = COPY $q1
+
+    %a:_(<8 x s32>) = G_CONCAT_VECTORS %p1:_(<4 x s32>), %p2:_(<4 x s32>)
+
+    %x:_(<4 x s32>) = G_SHUFFLE_VECTOR %a:_(<8 x s32>), %a:_, shufflemask(0, 1, 2, 3)
+    %y:_(<4 x s32>) = G_SHUFFLE_VECTOR %a:_(<8 x s32>), %a:_, shufflemask(4, 5, 6, 7)
+
+    $q0 = COPY %x(<4 x s32>)
+    $q1 = COPY %y(<4 x s32>)
+    RET_ReallyLR implicit $q0
+...