[x86] Teach the x86 DAG combiner to form MOVSLDUP and MOVSHDUP

instructions when it finds an appropriate pattern.

These are lovely instructions, and its a shame to not use them. =] They
are fast, and can hand loads folded into their operands, etc.

I've also plumbed the comment shuffle decoding through the various
layers so that the test cases are printed nicely.

llvm-svn: 217758

4 files changed, 105 insertions, 30 deletions

diff --git a/llvm/lib/Target/X86/InstPrinter/X86InstComments.cpp b/llvm/lib/Target/X86/InstPrinter/X86InstComments.cpp
index 5deb6825148..cb120633a40 100644
--- a/llvm/lib/Target/X86/InstPrinter/X86InstComments.cpp
+++ b/llvm/lib/Target/X86/InstPrinter/X86InstComments.cpp
@@ -128,6 +128,42 @@ bool llvm::EmitAnyX86InstComments(const MCInst *MI, raw_ostream &OS,
     DecodeMOVHLPSMask(2, ShuffleMask);
     break;
 
+  case X86::MOVSLDUPrr:
+  case X86::VMOVSLDUPrr:
+    Src1Name = getRegName(MI->getOperand(1).getReg());
+    // FALL THROUGH.
+  case X86::MOVSLDUPrm:
+  case X86::VMOVSLDUPrm:
+    DestName = getRegName(MI->getOperand(0).getReg());
+    DecodeMOVSLDUPMask(MVT::v4f32, ShuffleMask);
+    break;
+
+  case X86::VMOVSHDUPYrr:
+    Src1Name = getRegName(MI->getOperand(1).getReg());
+    // FALL THROUGH.
+  case X86::VMOVSHDUPYrm:
+    DestName = getRegName(MI->getOperand(0).getReg());
+    DecodeMOVSHDUPMask(MVT::v8f32, ShuffleMask);
+    break;
+
+  case X86::VMOVSLDUPYrr:
+    Src1Name = getRegName(MI->getOperand(1).getReg());
+    // FALL THROUGH.
+  case X86::VMOVSLDUPYrm:
+    DestName = getRegName(MI->getOperand(0).getReg());
+    DecodeMOVSLDUPMask(MVT::v8f32, ShuffleMask);
+    break;
+
+  case X86::MOVSHDUPrr:
+  case X86::VMOVSHDUPrr:
+    Src1Name = getRegName(MI->getOperand(1).getReg());
+    // FALL THROUGH.
+  case X86::MOVSHDUPrm:
+  case X86::VMOVSHDUPrm:
+    DestName = getRegName(MI->getOperand(0).getReg());
+    DecodeMOVSHDUPMask(MVT::v4f32, ShuffleMask);
+    break;
+
   case X86::PALIGNR128rr:
   case X86::VPALIGNR128rr:
     Src1Name = getRegName(MI->getOperand(2).getReg());
diff --git a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp
index 5da63bef4d7..51d251e551c 100644
--- a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp
+++ b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp
@@ -63,6 +63,22 @@ void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
     ShuffleMask.push_back(NElts+i);
 }
 
+void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
+  unsigned NumElts = VT.getVectorNumElements();
+  for (int i = 0, e = NumElts / 2; i < e; ++i) {
+    ShuffleMask.push_back(2 * i);
+    ShuffleMask.push_back(2 * i);
+  }
+}
+
+void DecodeMOVSHDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
+  unsigned NumElts = VT.getVectorNumElements();
+  for (int i = 0, e = NumElts / 2; i < e; ++i) {
+    ShuffleMask.push_back(2 * i + 1);
+    ShuffleMask.push_back(2 * i + 1);
+  }
+}
+
 void DecodePALIGNRMask(MVT VT, unsigned Imm,
                        SmallVectorImpl<int> &ShuffleMask) {
   unsigned NumElts = VT.getVectorNumElements();
diff --git a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h
index 911e7047ee6..03a843e7b8d 100644
--- a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h
+++ b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h
@@ -38,6 +38,10 @@ void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask);
 // <0,2> or <0,1,4,5>
 void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask);
 
+void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask);
+
+void DecodeMOVSHDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask);
+
 void DecodePALIGNRMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask);
 
 void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask);
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a4e94551020..04f1fafa2e7 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5407,12 +5407,16 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT,
     DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
     if (Mask.empty()) return false;
     break;
+  case X86ISD::MOVSLDUP:
+    DecodeMOVSLDUPMask(VT, Mask);
+    break;
+  case X86ISD::MOVSHDUP:
+    DecodeMOVSHDUPMask(VT, Mask);
+    break;
   case X86ISD::MOVDDUP:
   case X86ISD::MOVLHPD:
   case X86ISD::MOVLPD:
   case X86ISD::MOVLPS:
-  case X86ISD::MOVSHDUP:
-  case X86ISD::MOVSLDUP:
     // Not yet implemented
     return false;
   default: llvm_unreachable("unknown target shuffle node");
@@ -19364,38 +19368,53 @@ static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
   // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
   // vectors because it can have a load folded into it that UNPCK cannot. This
   // doesn't preclude something switching to the shorter encoding post-RA.
-  if (FloatDomain && (Mask.equals(0, 0) || Mask.equals(1, 1))) {
-    bool Lo = Mask.equals(0, 0);
-    unsigned Shuffle;
-    MVT ShuffleVT;
-    // Check if we have SSE3 which will let us use MOVDDUP. That instruction
-    // is no slower than UNPCKLPD but has the option to fold the input operand
-    // into even an unaligned memory load.
-    if (Lo && Subtarget->hasSSE3()) {
-      Shuffle = X86ISD::MOVDDUP;
-      ShuffleVT = MVT::v2f64;
-    } else {
-      // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
-      // than the UNPCK variants.
-      Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
-      ShuffleVT = MVT::v4f32;
+  if (FloatDomain) {
+    if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
+      bool Lo = Mask.equals(0, 0);
+      unsigned Shuffle;
+      MVT ShuffleVT;
+      // Check if we have SSE3 which will let us use MOVDDUP. That instruction
+      // is no slower than UNPCKLPD but has the option to fold the input operand
+      // into even an unaligned memory load.
+      if (Lo && Subtarget->hasSSE3()) {
+        Shuffle = X86ISD::MOVDDUP;
+        ShuffleVT = MVT::v2f64;
+      } else {
+        // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
+        // than the UNPCK variants.
+        Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
+        ShuffleVT = MVT::v4f32;
+      }
+      if (Depth == 1 && Root->getOpcode() == Shuffle)
+        return false; // Nothing to do!
+      Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
+      DCI.AddToWorklist(Op.getNode());
+      if (Shuffle == X86ISD::MOVDDUP)
+        Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
+      else
+        Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
+      DCI.AddToWorklist(Op.getNode());
+      DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+                    /*AddTo*/ true);
+      return true;
     }
-    if (Depth == 1 && Root->getOpcode() == Shuffle)
-      return false; // Nothing to do!
-    Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
-    DCI.AddToWorklist(Op.getNode());
-    if (Shuffle == X86ISD::MOVDDUP)
+    if (Subtarget->hasSSE3() &&
+        (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
+      bool Lo = Mask.equals(0, 0, 2, 2);
+      unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
+      MVT ShuffleVT = MVT::v4f32;
+      if (Depth == 1 && Root->getOpcode() == Shuffle)
+        return false; // Nothing to do!
+      Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
+      DCI.AddToWorklist(Op.getNode());
       Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
-    else
-      Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
-    DCI.AddToWorklist(Op.getNode());
-    DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
-                  /*AddTo*/ true);
-    return true;
+      DCI.AddToWorklist(Op.getNode());
+      DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+                    /*AddTo*/ true);
+      return true;
+    }
   }
 
-  // FIXME: We should match UNPCKLPS and UNPCKHPS here.
-
   // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
   // variants as none of these have single-instruction variants that are
   // superior to the UNPCK formulation.