[X86][SSE] Add support for target shuffle combining to INSERTPS

llvm-svn: 274990

3 files changed, 51 insertions, 18 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 4e3a85a70c6..83e910d4f14 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -8677,16 +8677,14 @@ static SDValue lowerVectorShuffleAsBroadcast(const SDLoc &DL, MVT VT,
 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
 // perform INSERTPS if a single V1 element is out of place and all V2
 // elements are zeroable.
-static SDValue lowerVectorShuffleAsInsertPS(const SDLoc &DL,
-                                            SDValue V1, SDValue V2,
-                                            ArrayRef<int> Mask,
-                                            SelectionDAG &DAG) {
+static bool matchVectorShuffleAsInsertPS(SDValue &V1, SDValue &V2,
+                                         unsigned &InsertPSMask,
+                                         const SmallBitVector &Zeroable,
+                                         ArrayRef<int> Mask,
+                                         SelectionDAG &DAG) {
   assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
   assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
   assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
-
-  SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
-
   unsigned ZMask = 0;
   int V1DstIndex = -1;
   int V2DstIndex = -1;
@@ -8707,7 +8705,7 @@ static SDValue lowerVectorShuffleAsInsertPS(const SDLoc &DL,
 
     // We can only insert a single non-zeroable element.
     if (V1DstIndex >= 0 || V2DstIndex >= 0)
-      return SDValue();
+      return false;
 
     if (Mask[i] < 4) {
       // V1 input out of place for insertion.
@@ -8720,7 +8718,7 @@ static SDValue lowerVectorShuffleAsInsertPS(const SDLoc &DL,
 
   // Don't bother if we have no (non-zeroable) element for insertion.
   if (V1DstIndex < 0 && V2DstIndex < 0)
-    return SDValue();
+    return false;
 
   // Determine element insertion src/dst indices. The src index is from the
   // start of the inserted vector, not the start of the concatenated vector.
@@ -8740,8 +8738,21 @@ static SDValue lowerVectorShuffleAsInsertPS(const SDLoc &DL,
   if (!V1UsedInPlace)
     V1 = DAG.getUNDEF(MVT::v4f32);
 
-  unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
+  // Insert the V2 element into the desired position.
+  InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
   assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
+  return true;
+}
+
+static SDValue lowerVectorShuffleAsInsertPS(const SDLoc &DL, SDValue V1,
+                                            SDValue V2, ArrayRef<int> Mask,
+                                            SelectionDAG &DAG) {
+  SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+
+  // Attempt to match the insertps pattern.
+  unsigned InsertPSMask;
+  if (!matchVectorShuffleAsInsertPS(V1, V2, InsertPSMask, Zeroable, Mask, DAG))
+    return SDValue();
 
   // Insert the V2 element into the desired position.
   return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
@@ -25081,6 +25092,33 @@ static bool combineX86ShuffleChain(SDValue Input, SDValue Root,
     }
   }
 
+  // Attempt to combine to INSERTPS.
+  if (Subtarget.hasSSE41() && NumMaskElts == 4 &&
+      (VT == MVT::v2f64 || VT == MVT::v4f32)) {
+    SmallBitVector Zeroable(4, false);
+    for (unsigned i = 0; i != NumMaskElts; ++i)
+      if (Mask[i] < 0)
+        Zeroable[i] = true;
+
+    unsigned InsertPSMask;
+    SDValue V1 = Input, V2 = Input;
+    if (Zeroable.any() && matchVectorShuffleAsInsertPS(V1, V2, InsertPSMask,
+                                                       Zeroable, Mask, DAG)) {
+      if (Depth == 1 && Root.getOpcode() == X86ISD::INSERTPS)
+        return false; // Nothing to do!
+      V1 = DAG.getBitcast(MVT::v4f32, V1);
+      DCI.AddToWorklist(V1.getNode());
+      V2 = DAG.getBitcast(MVT::v4f32, V2);
+      DCI.AddToWorklist(V2.getNode());
+      Res = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
+                        DAG.getConstant(InsertPSMask, DL, MVT::i8));
+      DCI.AddToWorklist(Res.getNode());
+      DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res),
+                    /*AddTo*/ true);
+      return true;
+    }
+  }
+
   // Don't try to re-form single instruction chains under any circumstances now
   // that we've done encoding canonicalization for them.
   if (Depth < 2)
diff --git a/llvm/test/CodeGen/X86/insertps-combine.ll b/llvm/test/CodeGen/X86/insertps-combine.ll
index 01d859e8d33..b21fdec624b 100644
--- a/llvm/test/CodeGen/X86/insertps-combine.ll
+++ b/llvm/test/CodeGen/X86/insertps-combine.ll
@@ -60,17 +60,12 @@ define <4 x float> @shuffle_v4f32_0z24(<4 x float> %xyzw, <4 x float> %abcd) {
 define <4 x float> @shuffle_v4f32_0zz0(float %a) {
 ; SSE-LABEL: shuffle_v4f32_0zz0:
 ; SSE:       # BB#0:
-; SSE-NEXT:    xorps %xmm1, %xmm1
-; SSE-NEXT:    blendps {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1,1,0]
-; SSE-NEXT:    movaps %xmm1, %xmm0
+; SSE-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],zero,zero,xmm0[0]
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: shuffle_v4f32_0zz0:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
-; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,0]
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],zero,zero,xmm0[0]
 ; AVX-NEXT:    retq
   %vecinit = insertelement <4 x float> undef, float %a, i32 0
   %vecinit1 = insertelement <4 x float> %vecinit, float 0.000000e+00, i32 1
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll
index ffae771da46..ea6535d390e 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll
@@ -190,7 +190,7 @@ define <8 x float> @combine_vpermilvar_8f32_4stage(<8 x float> %a0) {
 define <4 x float> @combine_vpermilvar_4f32_as_insertps(<4 x float> %a0) {
 ; ALL-LABEL: combine_vpermilvar_4f32_as_insertps:
 ; ALL:       # BB#0:
-; ALL-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[4,5,6,7],zero,zero,zero,zero,xmm0[8,9,10,11],zero,zero,zero,zero
+; ALL-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[1],zero,xmm0[2],zero
 ; ALL-NEXT:    retq
   %1 = call <4 x float> @llvm.x86.avx.vpermilvar.ps(<4 x float> %a0, <4 x i32> <i32 3, i32 2, i32 1, i32 0>)
   %2 = shufflevector <4 x float> %1, <4 x float> zeroinitializer, <4 x i32> <i32 2, i32 4, i32 1, i32 4>