Synthesize SSSE3/AVX 128-bit horizontal integer add/sub instructions from add/sub of appropriate shuffle vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144989 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 62 insertions, 3 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 2003e71f16..9585c09ad9 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -14670,7 +14670,24 @@ static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
                      DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
 }
 
-static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
+/// PerformADDCombine - Do target-specific dag combines on integer adds.
+static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
+                                 const X86Subtarget *Subtarget) {
+  EVT VT = N->getValueType(0);
+  SDValue Op0 = N->getOperand(0);
+  SDValue Op1 = N->getOperand(1);
+
+  // Try to synthesize horizontal adds from adds of shuffles.
+  if ((Subtarget->hasSSSE3() || Subtarget->hasAVX()) &&
+      (VT == MVT::v8i16 || VT == MVT::v4i32) &&
+      isHorizontalBinOp(Op0, Op1, true))
+    return DAG.getNode(X86ISD::HADD, N->getDebugLoc(), VT, Op0, Op1);
+
+  return OptimizeConditionalInDecrement(N, DAG);
+}
+
+static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
+                                 const X86Subtarget *Subtarget) {
   SDValue Op0 = N->getOperand(0);
   SDValue Op1 = N->getOperand(1);
 
@@ -14692,6 +14709,13 @@ static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
     }
   }
 
+  // Try to synthesize horizontal adds from adds of shuffles.
+  EVT VT = N->getValueType(0);
+  if ((Subtarget->hasSSSE3() || Subtarget->hasAVX()) &&
+      (VT == MVT::v8i16 || VT == MVT::v4i32) &&
+      isHorizontalBinOp(Op0, Op1, false))
+    return DAG.getNode(X86ISD::HSUB, N->getDebugLoc(), VT, Op0, Op1);
+
   return OptimizeConditionalInDecrement(N, DAG);
 }
 
@@ -14705,8 +14729,8 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::VSELECT:
   case ISD::SELECT:         return PerformSELECTCombine(N, DAG, Subtarget);
   case X86ISD::CMOV:        return PerformCMOVCombine(N, DAG, DCI);
-  case ISD::ADD:            return OptimizeConditionalInDecrement(N, DAG);
-  case ISD::SUB:            return PerformSubCombine(N, DAG);
+  case ISD::ADD:            return PerformAddCombine(N, DAG, Subtarget);
+  case ISD::SUB:            return PerformSubCombine(N, DAG, Subtarget);
   case X86ISD::ADC:         return PerformADCCombine(N, DAG, DCI);
   case ISD::MUL:            return PerformMulCombine(N, DAG, DCI);
   case ISD::SHL:
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index c95c45169b..5a407012c0 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -178,6 +178,12 @@ namespace llvm {
       /// BLEND family of opcodes
       BLENDV,
 
+      /// HADD - Integer horizontal add.
+      HADD,
+
+      /// HSUB - Integer horizontal sub.
+      HSUB,
+
       /// FHADD - Floating point horizontal add.
       FHADD,
 
diff --git a/lib/Target/X86/X86InstrFragmentsSIMD.td b/lib/Target/X86/X86InstrFragmentsSIMD.td
index 62ee94fd7c..c91e2dfefe 100644
--- a/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -41,6 +41,8 @@ def X86fsrl    : SDNode<"X86ISD::FSRL",      SDTX86FPShiftOp>;
 def X86fgetsign: SDNode<"X86ISD::FGETSIGNx86",SDTFPToIntOp>;
 def X86fhadd   : SDNode<"X86ISD::FHADD",     SDTFPBinOp>;
 def X86fhsub   : SDNode<"X86ISD::FHSUB",     SDTFPBinOp>;
+def X86hadd    : SDNode<"X86ISD::HADD",      SDTIntBinOp>;
+def X86hsub    : SDNode<"X86ISD::HSUB",      SDTIntBinOp>;
 def X86comi    : SDNode<"X86ISD::COMI",      SDTX86CmpTest>;
 def X86ucomi   : SDNode<"X86ISD::UCOMI",     SDTX86CmpTest>;
 def X86cmpss   : SDNode<"X86ISD::FSETCCss",    SDTX86Cmpss>;
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 27ad550056..c365fad73e 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -5369,6 +5369,15 @@ let Predicates = [HasSSSE3] in {
             (PSIGNWrr128 VR128:$src1, VR128:$src2)>;
   def : Pat<(v4i32 (X86psign VR128:$src1, VR128:$src2)),
             (PSIGNDrr128 VR128:$src1, VR128:$src2)>;
+
+  def : Pat<(v8i16 (X86hadd VR128:$src1, VR128:$src2)),
+            (PHADDWrr128 VR128:$src1, VR128:$src2)>;
+  def : Pat<(v4i32 (X86hadd VR128:$src1, VR128:$src2)),
+            (PHADDDrr128 VR128:$src1, VR128:$src2)>;
+  def : Pat<(v8i16 (X86hsub VR128:$src1, VR128:$src2)),
+            (PHSUBWrr128 VR128:$src1, VR128:$src2)>;
+  def : Pat<(v4i32 (X86hsub VR128:$src1, VR128:$src2)),
+            (PHSUBDrr128 VR128:$src1, VR128:$src2)>;
 }
 
 let Predicates = [HasAVX] in {
@@ -5383,6 +5392,15 @@ let Predicates = [HasAVX] in {
             (VPSIGNWrr128 VR128:$src1, VR128:$src2)>;
   def : Pat<(v4i32 (X86psign VR128:$src1, VR128:$src2)),
             (VPSIGNDrr128 VR128:$src1, VR128:$src2)>;
+
+  def : Pat<(v8i16 (X86hadd VR128:$src1, VR128:$src2)),
+            (VPHADDWrr128 VR128:$src1, VR128:$src2)>;
+  def : Pat<(v4i32 (X86hadd VR128:$src1, VR128:$src2)),
+            (VPHADDDrr128 VR128:$src1, VR128:$src2)>;
+  def : Pat<(v8i16 (X86hsub VR128:$src1, VR128:$src2)),
+            (VPHSUBWrr128 VR128:$src1, VR128:$src2)>;
+  def : Pat<(v4i32 (X86hsub VR128:$src1, VR128:$src2)),
+            (VPHSUBDrr128 VR128:$src1, VR128:$src2)>;
 }
 
 let Predicates = [HasAVX2] in {
@@ -5392,6 +5410,15 @@ let Predicates = [HasAVX2] in {
             (VPSIGNWrr256 VR256:$src1, VR256:$src2)>;
   def : Pat<(v8i32 (X86psign VR256:$src1, VR256:$src2)),
             (VPSIGNDrr256 VR256:$src1, VR256:$src2)>;
+
+  def : Pat<(v16i16 (X86hadd VR256:$src1, VR256:$src2)),
+            (VPHADDWrr256 VR256:$src1, VR256:$src2)>;
+  def : Pat<(v8i32 (X86hadd VR256:$src1, VR256:$src2)),
+            (VPHADDDrr256 VR256:$src1, VR256:$src2)>;
+  def : Pat<(v16i16 (X86hsub VR256:$src1, VR256:$src2)),
+            (VPHSUBWrr256 VR256:$src1, VR256:$src2)>;
+  def : Pat<(v8i32 (X86hsub VR256:$src1, VR256:$src2)),
+            (VPHSUBDrr256 VR256:$src1, VR256:$src2)>;
 }
 
 //===---------------------------------------------------------------------===//