AVX-512: Fixed a bug in VPERMT2* intrinsic.

[oota-llvm.git] / lib / Target / X86 / X86ISelLowering.cpp

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 26e69507617ca5bd658ed140b2ce6b2688f4b5f5..40c6bce6978e512be48004ba72c0969a7ee52f84 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -16377,7 +16377,27 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget
                                   Mask, PassThru, Subtarget, DAG);
     }
     case VPERM_3OP_MASKZ:
-    case VPERM_3OP_MASK:
+    case VPERM_3OP_MASK:{
+      // Src2 is the PassThru
+      SDValue Src1 = Op.getOperand(1);
+      SDValue Src2 = Op.getOperand(2);
+      SDValue Src3 = Op.getOperand(3);
+      SDValue Mask = Op.getOperand(4);
+      MVT VT = Op.getSimpleValueType();
+      SDValue PassThru = SDValue();
+
+      // set PassThru element
+      if (IntrData->Type == VPERM_3OP_MASKZ)
+        PassThru = getZeroVector(VT, Subtarget, DAG, dl);
+      else
+        PassThru = Src2;
+
+      // Swap Src1 and Src2 in the node creation
+      return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
+                                              dl, Op.getValueType(),
+                                              Src2, Src1, Src3),
+                                  Mask, PassThru, Subtarget, DAG);
+    }
     case FMA_OP_MASK3:
     case FMA_OP_MASKZ:
     case FMA_OP_MASK: {
@@ -16389,7 +16409,7 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget
       SDValue PassThru = SDValue();
 
       // set PassThru element
-      if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
+      if (IntrData->Type == FMA_OP_MASKZ)
         PassThru = getZeroVector(VT, Subtarget, DAG, dl);
       else if (IntrData->Type == FMA_OP_MASK3)
         PassThru = Src3;