[AArch64][GlobalISel] Combine G_UNMERGE(G_DUPLANE16) -> G_DUPLANE16

We will generate G_UNMERGE(G_DUPLANE16) due to the legalization of shuffle
vector splats with mismatching vector sizes. The G_DUPLANE intrinsics can
handle different vector sizes (128bit and 64bit output, for example), and we
can combine away the unmerge.

Author: davemgreen

llvm/lib/Target/AArch64/AArch64Combine.td

@@ -172,6 +172,20 @@ def form_duplane : GICombineRule <
   (apply [{ applyDupLane(*${root}, MRI, B, ${matchinfo}); }])
 >;
 
+// Clean up G_UNMERGE(G_DUPLANE16) -> G_DUPLANE16
+class unmerge_duplane<Instruction Op> : GICombineRule <
+  (defs root:$root),
+  (match (Op $a, $src, $c),
+         (G_UNMERGE_VALUES $d1, $d2, $a):$root,
+         [{ return MRI.getType(${d1}.getReg()).getSizeInBits() == 64; }]),
+  (apply (GIReplaceReg $d2, $d1), (Op $d1, $src, $c))
+>;
+def unmerge_duplane8 : unmerge_duplane<G_DUPLANE8>;
+def unmerge_duplane16 : unmerge_duplane<G_DUPLANE16>;
+def unmerge_duplane32 : unmerge_duplane<G_DUPLANE32>;
+def unmerge_duplane64 : unmerge_duplane<G_DUPLANE64>;
+def unmerge_duplanes : GICombineGroup<[unmerge_duplane8, unmerge_duplane16, unmerge_duplane32, unmerge_duplane64]>;
+
 def shuffle_vector_lowering : GICombineGroup<[dup, rev, ext, zip, uzp, trn, fullrev,
                                               form_duplane, shuf_to_ins]>;
 
@@ -325,7 +339,8 @@ def AArch64PostLegalizerLowering
                         lower_vector_fcmp, form_truncstore,
                         vector_sext_inreg_to_shift,
                         unmerge_ext_to_unmerge, lower_mulv2s64,
-                        vector_unmerge_lowering, insertelt_nonconst]> {
+                        vector_unmerge_lowering, insertelt_nonconst,
+                        unmerge_duplanes]> {
 }
 
 // Post-legalization combines which are primarily optimizations.

llvm/test/CodeGen/AArch64/arm64-dup.ll

@@ -422,16 +422,10 @@ define <4 x i16> @test_build_illegal(<4 x i32> %in) {
 ; SelectionDAGBuilder here. We then added a DUPLANE on top of that, preventing
 ; the formation of an indexed-by-7 MLS.
 define <4 x i16> @test_high_splat(<4 x i16> %a, <4 x i16> %b, <8 x i16> %v) #0 {
-; CHECK-SD-LABEL: test_high_splat:
-; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    mls.4h v0, v1, v2[7]
-; CHECK-SD-NEXT:    ret
-;
-; CHECK-GI-LABEL: test_high_splat:
-; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    dup.8h v2, v2[7]
-; CHECK-GI-NEXT:    mls.4h v0, v2, v1
-; CHECK-GI-NEXT:    ret
+; CHECK-LABEL: test_high_splat:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    mls.4h v0, v1, v2[7]
+; CHECK-NEXT:    ret
 entry:
   %shuffle = shufflevector <8 x i16> %v, <8 x i16> undef, <4 x i32> <i32 7, i32 7, i32 7, i32 7>
   %mul = mul <4 x i16> %shuffle, %b