diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -375,6 +375,12 @@
   "True 16-bit operand instructions"
 >;
 
+def FeatureRealTrue16Insts : SubtargetFeature<"real-true16",
+  "EnableRealTrue16Insts",
+  "true",
+  "Use true 16-bit registers"
+>;
+
 def FeatureVOP3P : SubtargetFeature<"vop3p",
   "HasVOP3PInsts",
   "true",
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -49,6 +49,7 @@
   bool GCN3Encoding = false;
   bool Has16BitInsts = false;
   bool HasTrue16BitInsts = false;
+  bool EnableRealTrue16Insts = false;
   bool HasMadMixInsts = false;
   bool HasMadMacF32Insts = false;
   bool HasDsSrc2Insts = false;
@@ -153,8 +154,17 @@
     return Has16BitInsts;
   }
 
+  /// Return true if the subtarget supports True16 instructions.
   bool hasTrue16BitInsts() const { return HasTrue16BitInsts; }
 
+  /// Return true if real (non-fake) variants of True16 instructions using
+  /// 16-bit registers should be code-generated. Fake True16 instructions are
+  /// identical to non-fake ones except that they take 32-bit registers as
+  /// operands and always use their low halves.
+  // TODO: Remove and use hasTrue16BitInsts() instead once True16 is fully
+  // supported and the support for fake True16 instructions is removed.
+  bool useRealTrue16Insts() const;
+
   bool hasMadMixInsts() const {
     return HasMadMixInsts;
   }
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -167,6 +167,10 @@
 
 AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT) : TargetTriple(TT) {}
 
+bool AMDGPUSubtarget::useRealTrue16Insts() const {
+  return hasTrue16BitInsts() && EnableRealTrue16Insts;
+}
+
 GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
                            const GCNTargetMachine &TM)
     : // clang-format off
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -148,8 +148,13 @@
   addRegisterClass(MVT::v16f64, TRI->getVGPRClassForBitWidth(1024));
 
   if (Subtarget->has16BitInsts()) {
-    addRegisterClass(MVT::i16, &AMDGPU::SReg_32RegClass);
-    addRegisterClass(MVT::f16, &AMDGPU::SReg_32RegClass);
+    if (Subtarget->useRealTrue16Insts()) {
+      addRegisterClass(MVT::i16, &AMDGPU::VGPR_16RegClass);
+      addRegisterClass(MVT::f16, &AMDGPU::VGPR_16RegClass);
+    } else {
+      addRegisterClass(MVT::i16, &AMDGPU::SReg_32RegClass);
+      addRegisterClass(MVT::f16, &AMDGPU::SReg_32RegClass);
+    }
 
     // Unless there are also VOP3P operations, not operations are really legal.
     addRegisterClass(MVT::v2i16, &AMDGPU::SReg_32RegClass);