[DAG] Extend visitSCALAR_TO_VECTOR optimization to truncated vector.

Summary:
Allow SCALAR_TO_VECTOR of EXTRACT_VECTOR_ELT to reduce to
EXTRACT_SUBVECTOR of vector shuffle when output is smaller. Marginally
improves vector shuffle computations.

Reviewers: efriedma, RKSimon, spatel

Subscribers: javed.absar, llvm-commits

Differential Revision: https://reviews.llvm.org/D35566

llvm-svn: 309680

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 850558629170910cd7812f1aa90e5aa2b8f9037c..5c9db976a42eb5d0fc719c81d746fe5f02fe4151 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -15710,23 +15710,38 @@ SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
   EVT VT = N->getValueType(0);
 
   // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
-  // with a VECTOR_SHUFFLE.
+  // with a VECTOR_SHUFFLE and possible truncate.
   if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
     SDValue InVec = InVal->getOperand(0);
     SDValue EltNo = InVal->getOperand(1);
-
-    // FIXME: We could support implicit truncation if the shuffle can be
-    // scaled to a smaller vector scalar type.
+    auto InVecT = InVec.getValueType();
     ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
-    if (C0 && VT == InVec.getValueType() &&
-        VT.getScalarType() == InVal.getValueType()) {
-      SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
+
+    if (C0) {
+      SmallVector<int, 8> NewMask(InVecT.getVectorNumElements(), -1);
       int Elt = C0->getZExtValue();
       NewMask[0] = Elt;
-
-      if (TLI.isShuffleMaskLegal(NewMask, VT))
-        return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
-                                    NewMask);
+      SDValue Val;
+      if (VT.getVectorNumElements() <= InVecT.getVectorNumElements() &&
+          TLI.isShuffleMaskLegal(NewMask, VT)) {
+        Val = DAG.getVectorShuffle(InVecT, SDLoc(N), InVec,
+                                   DAG.getUNDEF(InVecT), NewMask);
+        // If the initial vector is the correct size this shuffle is a
+        // valid result.
+        if (VT == InVecT)
+          return Val;
+        // If not we must truncate the vector.
+        if (VT.getVectorNumElements() != InVecT.getVectorNumElements()) {
+          MVT IdxTy = TLI.getVectorIdxTy(DAG.getDataLayout());
+          SDValue ZeroIdx = DAG.getConstant(0, SDLoc(N), IdxTy);
+          EVT SubVT =
+              EVT::getVectorVT(*DAG.getContext(), InVecT.getVectorElementType(),
+                               VT.getVectorNumElements());
+          Val = DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), SubVT, Val,
+                            ZeroIdx);
+          return Val;
+        }
+      }
     }
   }
 

diff --git a/llvm/test/CodeGen/AArch64/arm64-neon-copy.ll b/llvm/test/CodeGen/AArch64/arm64-neon-copy.ll
index 16834b7e315b19654140e2a43b38ae71594fc961..2585676e1bd290284e9b0de116b04c89884aa7af 100644 (file)
--- a/llvm/test/CodeGen/AArch64/arm64-neon-copy.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-neon-copy.ll
@@ -188,7 +188,7 @@ define <2 x float> @ins4f2(<4 x float> %tmp1, <2 x float> %tmp2) {
 
 define <1 x double> @ins2f1(<2 x double> %tmp1, <1 x double> %tmp2) {
 ; CHECK-LABEL: ins2f1:
-; CHECK: mov {{d[0-9]+}}, {{v[0-9]+}}.d[1]
+; CHECK: dup {{v[0-9]+}}.2d, {{v[0-9]+}}.d[1]
   %tmp3 = extractelement <2 x double> %tmp1, i32 1
   %tmp4 = insertelement <1 x double> %tmp2, double %tmp3, i32 0
   ret <1 x double> %tmp4

diff --git a/llvm/test/CodeGen/AArch64/neon-scalar-copy.ll b/llvm/test/CodeGen/AArch64/neon-scalar-copy.ll
index 3f770600ac5929fe2f7a66cf03de1c91e843af4d..2384e485fd73ac44672e8fdff0b949f9c2c1bd35 100644 (file)
--- a/llvm/test/CodeGen/AArch64/neon-scalar-copy.ll
+++ b/llvm/test/CodeGen/AArch64/neon-scalar-copy.ll
@@ -79,8 +79,7 @@ define half @test_dup_hv8H_0(<8 x half> %v) #0 {
 
 define <1 x i8> @test_vector_dup_bv16B(<16 x i8> %v1) #0 {
  ; CHECK-LABEL: test_vector_dup_bv16B:
- ; CHECK-NEXT: umov [[W:w[0-9]+]], v0.b[14]
- ; CHECK-NEXT: fmov s0, [[W]]
+ ; CHECK-NEXT: dup v0.16b, v0.b[14]
  ; CHECK-NEXT: ret
  %shuffle.i = shufflevector <16 x i8> %v1, <16 x i8> undef, <1 x i32> <i32 14> 
  ret <1 x i8> %shuffle.i
@@ -96,8 +95,7 @@ define <1 x i8> @test_vector_dup_bv8B(<8 x i8> %v1) #0 {
 
 define <1 x i16> @test_vector_dup_hv8H(<8 x i16> %v1) #0 {
  ; CHECK-LABEL: test_vector_dup_hv8H:
- ; CHECK-NEXT: umov [[W:w[0-9]+]], v0.h[7]
- ; CHECK-NEXT: fmov s0, [[W]]
+ ; CHECK-NEXT: dup v0.8h, v0.h[7]
  ; CHECK-NEXT: ret
  %shuffle.i = shufflevector <8 x i16> %v1, <8 x i16> undef, <1 x i32> <i32 7> 
  ret <1 x i16> %shuffle.i
@@ -113,8 +111,7 @@ define <1 x i16> @test_vector_dup_hv4H(<4 x i16> %v1) #0 {
 
 define <1 x i32> @test_vector_dup_sv4S(<4 x i32> %v1) #0 {
  ; CHECK-LABEL: test_vector_dup_sv4S:
- ; CHECK-NEXT: mov  [[W:w[0-9]+]], v0.s[3]
- ; CHECK-NEXT: fmov s0, [[W]]
+ ; CHECK-NEXT: dup v0.4s, v0.s[3]
  ; CHECK-NEXT: ret
  %shuffle = shufflevector <4 x i32> %v1, <4 x i32> undef, <1 x i32> <i32 3> 
  ret <1 x i32> %shuffle
@@ -138,7 +135,7 @@ define <1 x i64> @test_vector_dup_dv2D(<2 x i64> %v1) #0 {
 
 define <1 x i64> @test_vector_copy_dup_dv2D(<1 x i64> %a, <2 x i64> %c) #0 {
   ; CHECK-LABEL: test_vector_copy_dup_dv2D:
-  ; CHECK-NEXT: {{dup|mov}} {{d[0-9]+}}, {{v[0-9]+}}.d[1]
+  ; CHECK-NEXT: dup v0.2d, v1.d[1]
   ; CHECK-NEXT: ret
   %vget_lane = extractelement <2 x i64> %c, i32 1
   %vset_lane = insertelement <1 x i64> undef, i64 %vget_lane, i32 0