[X86][MMX] Handle i32->mmx conversion using movd

Implement a BITCAST dag combine to transform i32->mmx conversion patterns
into a X86 specific node (MMX_MOVW2D) and guarantee that moves between
i32 and x86mmx are better handled, i.e., don't use store-load to do the
conversion..

llvm-svn: 228293

4 files changed, 38 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 17d0fce54f8..eff03f030c5 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -1675,6 +1675,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
   // We have target-specific dag combine patterns for the following nodes:
   setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
   setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
+  setTargetDAGCombine(ISD::BITCAST);
   setTargetDAGCombine(ISD::VSELECT);
   setTargetDAGCombine(ISD::SELECT);
   setTargetDAGCombine(ISD::SHL);
@@ -22986,6 +22987,25 @@ static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
                      EltNo);
 }
 
+/// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
+/// special and don't usually play with other vector types, it's better to
+/// handle them early to be sure we emit efficient code by avoiding
+/// store-load conversions.
+static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
+  if (N->getValueType(0) != MVT::x86mmx ||
+      N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
+      N->getOperand(0)->getValueType(0) != MVT::v2i32)
+    return SDValue();
+
+  SDValue V = N->getOperand(0);
+  ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
+  if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
+    return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
+                       N->getValueType(0), V.getOperand(0));
+
+  return SDValue();
+}
+
 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
 /// generation and convert it from being a bunch of shuffles and extracts
 /// into a somewhat faster sequence. For i686, the best sequence is apparently
@@ -26129,6 +26149,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::SELECT:
   case X86ISD::SHRUNKBLEND:
     return PerformSELECTCombine(N, DAG, DCI, Subtarget);
+  case ISD::BITCAST:        return PerformBITCASTCombine(N, DAG);
   case X86ISD::CMOV:        return PerformCMOVCombine(N, DAG, DCI, Subtarget);
   case ISD::ADD:            return PerformAddCombine(N, DAG, Subtarget);
   case ISD::SUB:            return PerformSubCombine(N, DAG, Subtarget);
diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
index 346f98e6bca..d90dcc71451 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.h
+++ b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -158,6 +158,10 @@ namespace llvm {
       /// vector to a GPR.
       MMX_MOVD2W,
 
+      /// MMX_MOVW2D - Copies a GPR into the low 32-bit word of a MMX vector
+      /// and zero out the high word.
+      MMX_MOVW2D,
+
       /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
       /// i32, corresponds to X86::PEXTRB.
       PEXTRB,
diff --git a/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td b/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
index f65292e7602..fd0a4ba28e8 100644
--- a/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -18,6 +18,9 @@
 // Low word of MMX to GPR.
 def MMX_X86movd2w : SDNode<"X86ISD::MMX_MOVD2W", SDTypeProfile<1, 1,
                             [SDTCisVT<0, i32>, SDTCisVT<1, x86mmx>]>>;
+// GPR to low word of MMX.
+def MMX_X86movw2d : SDNode<"X86ISD::MMX_MOVW2D", SDTypeProfile<1, 1,
+                            [SDTCisVT<0, x86mmx>, SDTCisVT<1, i32>]>>;
 
 //===----------------------------------------------------------------------===//
 // MMX Pattern Fragments
diff --git a/llvm/lib/Target/X86/X86InstrMMX.td b/llvm/lib/Target/X86/X86InstrMMX.td
index b5262c5cdac..fbec0625418 100644
--- a/llvm/lib/Target/X86/X86InstrMMX.td
+++ b/llvm/lib/Target/X86/X86InstrMMX.td
@@ -229,6 +229,16 @@ def MMX_MOVD64rm : MMXI<0x6E, MRMSrcMem, (outs VR64:$dst), (ins i32mem:$src),
                         [(set VR64:$dst,
                         (x86mmx (scalar_to_vector (loadi32 addr:$src))))],
                         IIC_MMX_MOV_MM_RM>, Sched<[WriteLoad]>;
+
+let Predicates = [HasMMX] in {
+  let AddedComplexity = 15 in
+    def : Pat<(x86mmx (MMX_X86movw2d GR32:$src)),
+              (MMX_MOVD64rr GR32:$src)>;
+  let AddedComplexity = 20 in
+    def : Pat<(x86mmx (MMX_X86movw2d (loadi32 addr:$src))),
+              (MMX_MOVD64rm addr:$src)>;
+}
+
 let mayStore = 1 in
 def MMX_MOVD64mr : MMXI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR64:$src),
                         "movd\t{$src, $dst|$dst, $src}", [], IIC_MMX_MOV_MM_RM>,