Index: lib/Target/AMDGPU/SIDefines.h
===================================================================
--- lib/Target/AMDGPU/SIDefines.h
+++ lib/Target/AMDGPU/SIDefines.h
@@ -126,7 +126,10 @@
 
     /// Operand with 32-bit immediate that uses the constant bus.
     OPERAND_KIMM32,
-    OPERAND_KIMM16
+    OPERAND_KIMM16,
+
+    // Output register operand with 16-bit output.
+    OPERAND_REG_DEF16
   };
 }
 
Index: lib/Target/AMDGPU/SIInstrFormats.td
===================================================================
--- lib/Target/AMDGPU/SIInstrFormats.td
+++ lib/Target/AMDGPU/SIInstrFormats.td
@@ -187,6 +187,22 @@
 }
 
 class VOPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVOPDst">;
+class VOP3DstOperand <RegisterClass rc> : RegisterOperand <rc>;
+
+let OperandNamespace = "AMDGPU" in {
+def VOPDstOperand16 : VOPDstOperand<VGPR_32> {
+  let OperandType = "OPERAND_REG_DEF16";
+}
+
+def VOP3DstOperand16 : VOP3DstOperand<VGPR_32> {
+  let OperandType = "OPERAND_REG_DEF16";
+}
+
+}
+
+def VOPDstOperand32 : VOPDstOperand<VGPR_32>;
+def VOPDstOperand64 : VOPDstOperand<VReg_64>;
+def VOPDstOperand128 : VOPDstOperand<VReg_128>;
 
 class VINTRPe <bits<2> op> : Enc32 {
   bits<8> vdst;
Index: lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.cpp
+++ lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -4540,7 +4540,21 @@
     KnownZero = 0xffff0000;
     return;
   }
-  default:
+  default: {
+    // XXX - Are there any 16-bit output instructions with second defs?
+    // FIXME:
+    assert(Def->getOperand(0).getReg() == Op.getReg() &&
+           Def->getOperand(0).getSubReg() == Op.getSubReg());
+    if (Def->getDesc().OpInfo[0].OperandType == AMDGPU::OPERAND_REG_DEF16) {
+      // FIXME: This isn't true for all instructions on gfx9, where some new
+      // instructions default to leaving high bits intact and there is a control
+      // bit for old instructions to change zeroing behavior.
+
+      KnownZero = 0xffff0000;
+      return;
+    }
+
     return;
+    }
   }
 }
Index: lib/Target/AMDGPU/SIInstrInfo.td
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.td
+++ lib/Target/AMDGPU/SIInstrInfo.td
@@ -886,10 +886,19 @@
   RegisterOperand ret = !if(!eq(VT.Size, 32), VOPDstOperand<VGPR_32>,
                           !if(!eq(VT.Size, 128), VOPDstOperand<VReg_128>,
                             !if(!eq(VT.Size, 64), VOPDstOperand<VReg_64>,
-                              !if(!eq(VT.Size, 16), VOPDstOperand<VGPR_32>,
+                              !if(!eq(VT.Size, 16), VOPDstOperand16,
                               VOPDstOperand<SReg_64>)))); // else VT == i1
 }
 
+// Hack to avoid printing _e64 for instructions that must be VOP3.
+class getVALUDstForVT_VOP3<ValueType VT> {
+  RegisterOperand ret = !if(!eq(VT.Size, 32), VOP3DstOperand<VGPR_32>,
+                          !if(!eq(VT.Size, 128), VOP3DstOperand<VReg_128>,
+                            !if(!eq(VT.Size, 64), VOP3DstOperand<VReg_64>,
+                              !if(!eq(VT.Size, 16), VOP3DstOperand16,
+                              VOP3DstOperand<SReg_64>)))); // else VT == i1
+}
+
 // Returns the register class to use for the destination of VOP[12C]
 // instructions with SDWA extension
 class getSDWADstForVT<ValueType VT> {
Index: lib/Target/AMDGPU/VOP3Instructions.td
===================================================================
--- lib/Target/AMDGPU/VOP3Instructions.td
+++ lib/Target/AMDGPU/VOP3Instructions.td
@@ -121,10 +121,13 @@
 // only VOP instruction that implicitly reads VCC.
 let Asm64 = " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$clamp$omod" in {
 def VOP_F32_F32_F32_F32_VCC : VOPProfile<[f32, f32, f32, f32]> {
-  let Outs64 = (outs DstRC.RegClass:$vdst);
+  // FIXME: Hack to stop printing _e64
+  let DstRC = getVALUDstForVT_VOP3<f32>.ret;
 }
+
 def VOP_F64_F64_F64_F64_VCC : VOPProfile<[f64, f64, f64, f64]> {
-  let Outs64 = (outs DstRC.RegClass:$vdst);
+  // FIXME: Hack to stop printing _e64
+  let DstRC = getVALUDstForVT_VOP3<f64>.ret;
 }
 }
 
@@ -139,7 +142,7 @@
 
 class VOP3_Profile<VOPProfile P> : VOPProfile<P.ArgVT> {
   // FIXME: Hack to stop printing _e64
-  let Outs64 = (outs DstRC.RegClass:$vdst);
+  let DstRC = getVALUDstForVT_VOP3<DstVT>.ret;
   let Asm64 = " " # P.Asm64;
 }
 
@@ -158,17 +161,17 @@
 
 def VOP3b_F32_I1_F32_F32_F32 : VOP3b_Profile<f32> {
   // FIXME: Hack to stop printing _e64
-  let DstRC = RegisterOperand<VGPR_32>;
+  let DstRC = getVALUDstForVT_VOP3<f32>.ret;
 }
 
 def VOP3b_F64_I1_F64_F64_F64 : VOP3b_Profile<f64> {
   // FIXME: Hack to stop printing _e64
-  let DstRC = RegisterOperand<VReg_64>;
+  let DstRC = getVALUDstForVT_VOP3<f64>.ret;
 }
 
 def VOP3b_I64_I1_I32_I32_I64 : VOPProfile<[i64, i32, i32, i64]> {
   // FIXME: Hack to stop printing _e64
-  let DstRC = RegisterOperand<VReg_64>;
+  let DstRC = getVALUDstForVT_VOP3<i64>.ret;
 
   let Outs64 = (outs DstRC:$vdst, SReg_64:$sdst);
   let Asm64 = " $vdst, $sdst, $src0, $src1, $src2";