[AMDGPU] Add a16 feature to gfx10

Based on D72931

This adds a new feature called A16 which is enabled for gfx10.
gfx9 keeps the R128A16 feature so it can share all the instruction encodings
with gfx7/8.

Differential Revision: https://reviews.llvm.org/D73956

14 files changed, 102 insertions, 36 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 42b477e..1064803 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -360,7 +360,13 @@ def FeatureDPP8 : SubtargetFeature<"dpp8",
 def FeatureR128A16 : SubtargetFeature<"r128-a16",
   "HasR128A16",
   "true",
-  "Support 16 bit coordindates/gradients/lod/clamp/mip types on gfx9"
+  "Support gfx9-style A16 for 16-bit coordinates/gradients/lod/clamp/mip image operands, where a16 is aliased with r128"
+>;
+
+def FeatureGFX10A16 : SubtargetFeature<"a16",
+  "HasGFX10A16",
+  "true",
+  "Support gfx10-style A16 for 16-bit coordinates/gradients/lod/clamp/mip image operands"
 >;
 
 def FeatureNSAEncoding : SubtargetFeature<"nsa-encoding",
@@ -682,7 +688,8 @@ def FeatureGFX10 : GCNSubtargetFeatureGeneration<"GFX10",
    FeatureAddNoCarryInsts, FeatureFmaMixInsts, FeatureGFX8Insts,
    FeatureNoSdstCMPX, FeatureVscnt, FeatureRegisterBanking,
    FeatureVOP3Literal, FeatureDPP8,
-   FeatureNoDataDepHazard, FeaturePkFmacF16Inst, FeatureDoesNotSupportSRAMECC
+   FeatureNoDataDepHazard, FeaturePkFmacF16Inst, FeatureDoesNotSupportSRAMECC,
+   FeatureGFX10A16
   ]
 >;
 
@@ -1094,6 +1101,9 @@ def HasDPP8 : Predicate<"Subtarget->hasDPP8()">,
 def HasR128A16 : Predicate<"Subtarget->hasR128A16()">,
   AssemblerPredicate<"FeatureR128A16">;
 
+def HasGFX10A16 : Predicate<"Subtarget->hasGFX10A16()">,
+  AssemblerPredicate<"FeatureGFX10A16">;
+
 def HasDPP16 : Predicate<"Subtarget->hasDPP()">,
   AssemblerPredicate<"!FeatureGCN3Encoding,FeatureGFX10Insts,FeatureDPP">;
 
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index 75c4fba..31c06ce 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -241,6 +241,7 @@ GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     HasDPP(false),
     HasDPP8(false),
     HasR128A16(false),
+    HasGFX10A16(false),
     HasNSAEncoding(false),
     HasDLInsts(false),
     HasDot1Insts(false),
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 6f3ca01..13fa915 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -342,6 +342,7 @@ protected:
   bool HasDPP;
   bool HasDPP8;
   bool HasR128A16;
+  bool HasGFX10A16;
   bool HasNSAEncoding;
   bool HasDLInsts;
   bool HasDot1Insts;
@@ -992,6 +993,10 @@ public:
     return HasR128A16;
   }
 
+  bool hasGFX10A16() const {
+    return HasGFX10A16;
+  }
+
   bool hasOffset3fBug() const {
     return HasOffset3fBug;
   }
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index ea70307..8c1f63e 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -163,6 +163,7 @@ public:
     ImmTyUNorm,
     ImmTyDA,
     ImmTyR128A16,
+    ImmTyA16,
     ImmTyLWE,
     ImmTyExpTgt,
     ImmTyExpCompr,
@@ -315,6 +316,7 @@ public:
   bool isUNorm() const { return isImmTy(ImmTyUNorm); }
   bool isDA() const { return isImmTy(ImmTyDA); }
   bool isR128A16() const { return isImmTy(ImmTyR128A16); }
+  bool isGFX10A16() const { return isImmTy(ImmTyA16); }
   bool isLWE() const { return isImmTy(ImmTyLWE); }
   bool isOff() const { return isImmTy(ImmTyOff); }
   bool isExpTgt() const { return isImmTy(ImmTyExpTgt); }
@@ -847,6 +849,7 @@ public:
     case ImmTyUNorm: OS << "UNorm"; break;
     case ImmTyDA: OS << "DA"; break;
     case ImmTyR128A16: OS << "R128A16"; break;
+    case ImmTyA16: OS << "A16"; break;
     case ImmTyLWE: OS << "LWE"; break;
     case ImmTyOff: OS << "Off"; break;
     case ImmTyExpTgt: OS << "ExpTgt"; break;
@@ -1157,6 +1160,10 @@ public:
     return AMDGPU::hasPackedD16(getSTI());
   }
 
+  bool hasGFX10A16() const {
+    return AMDGPU::hasGFX10A16(getSTI());
+  }
+
   bool isSI() const {
     return AMDGPU::isSI(getSTI());
   }
@@ -4650,9 +4657,9 @@ AMDGPUAsmParser::parseNamedBit(const char *Name, OperandVector &Operands,
       case AsmToken::Identifier: {
         StringRef Tok = Parser.getTok().getString();
         if (Tok == Name) {
-          if (Tok == "r128" && isGFX9())
+          if (Tok == "r128" && !hasMIMG_R128())
             Error(S, "r128 modifier is not supported on this GPU");
-          if (Tok == "a16" && !isGFX9() && !isGFX10())
+          if (Tok == "a16" && !isGFX9() && !hasGFX10A16())
             Error(S, "a16 modifier is not supported on this GPU");
           Bit = 1;
           Parser.Lex();
@@ -4672,6 +4679,9 @@ AMDGPUAsmParser::parseNamedBit(const char *Name, OperandVector &Operands,
   if (!isGFX10() && ImmTy == AMDGPUOperand::ImmTyDLC)
     return MatchOperand_ParseFail;
 
+  if (isGFX9() && ImmTy == AMDGPUOperand::ImmTyA16)
+    ImmTy = AMDGPUOperand::ImmTyR128A16;
+
   Operands.push_back(AMDGPUOperand::CreateImm(this, Bit, S, ImmTy));
   return MatchOperand_Success;
 }
@@ -5987,6 +5997,8 @@ void AMDGPUAsmParser::cvtMIMG(MCInst &Inst, const OperandVector &Operands,
   addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTyGLC);
   addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTySLC);
   addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTyR128A16);
+  if (IsGFX10)
+    addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTyA16);
   addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTyTFE);
   addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTyLWE);
   if (!IsGFX10)
@@ -6096,7 +6108,7 @@ static const OptionalOperand AMDGPUOptionalOperandTable[] = {
   {"unorm",   AMDGPUOperand::ImmTyUNorm, true, nullptr},
   {"da",      AMDGPUOperand::ImmTyDA,    true, nullptr},
   {"r128",    AMDGPUOperand::ImmTyR128A16,  true, nullptr},
-  {"a16",     AMDGPUOperand::ImmTyR128A16,  true, nullptr},
+  {"a16",     AMDGPUOperand::ImmTyA16,  true, nullptr},
   {"lwe",     AMDGPUOperand::ImmTyLWE,   true, nullptr},
   {"d16",     AMDGPUOperand::ImmTyD16,   true, nullptr},
   {"dmask",   AMDGPUOperand::ImmTyDMask, false, nullptr},
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
index f65dc25..55146b5 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
@@ -244,6 +244,11 @@ void AMDGPUInstPrinter::printR128A16(const MCInst *MI, unsigned OpNo,
     printNamedBit(MI, OpNo, O, "r128");
 }
 
+void AMDGPUInstPrinter::printGFX10A16(const MCInst *MI, unsigned OpNo,
+                                  const MCSubtargetInfo &STI, raw_ostream &O) {
+  printNamedBit(MI, OpNo, O, "a16");
+}
+
 void AMDGPUInstPrinter::printLWE(const MCInst *MI, unsigned OpNo,
                                  const MCSubtargetInfo &STI, raw_ostream &O) {
   printNamedBit(MI, OpNo, O, "lwe");
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
index ba53003..83f99f1 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
@@ -86,6 +86,8 @@ private:
                raw_ostream &O);
   void printR128A16(const MCInst *MI, unsigned OpNo, const MCSubtargetInfo &STI,
                  raw_ostream &O);
+  void printGFX10A16(const MCInst *MI, unsigned OpNo, const MCSubtargetInfo &STI,
+                raw_ostream &O);
   void printLWE(const MCInst *MI, unsigned OpNo,
                 const MCSubtargetInfo &STI, raw_ostream &O);
   void printD16(const MCInst *MI, unsigned OpNo,
diff --git a/llvm/lib/Target/AMDGPU/MIMGInstructions.td b/llvm/lib/Target/AMDGPU/MIMGInstructions.td
index 4006a62..8a43942 100644
--- a/llvm/lib/Target/AMDGPU/MIMGInstructions.td
+++ b/llvm/lib/Target/AMDGPU/MIMGInstructions.td
@@ -238,9 +238,9 @@ class MIMG_NoSampler_gfx10<int op, string opcode,
   : MIMG_gfx10<op, (outs DataRC:$vdata), dns> {
   let InOperandList = !con((ins AddrRC:$vaddr0, SReg_256:$srsrc, DMask:$dmask,
                                 Dim:$dim, UNorm:$unorm, DLC:$dlc, GLC:$glc,
-                                SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe),
+                                SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe),
                            !if(BaseOpcode.HasD16, (ins D16:$d16), (ins)));
-  let AsmString = opcode#" $vdata, $vaddr0, $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$tfe$lwe"
+  let AsmString = opcode#" $vdata, $vaddr0, $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$a16$tfe$lwe"
                     #!if(BaseOpcode.HasD16, "$d16", "");
 }
 
@@ -251,9 +251,9 @@ class MIMG_NoSampler_nsa_gfx10<int op, string opcode,
   let InOperandList = !con(AddrIns,
                            (ins SReg_256:$srsrc, DMask:$dmask,
                                 Dim:$dim, UNorm:$unorm, DLC:$dlc, GLC:$glc,
-                                SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe),
+                                SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe),
                            !if(BaseOpcode.HasD16, (ins D16:$d16), (ins)));
-  let AsmString = opcode#" $vdata, "#AddrAsm#", $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$tfe$lwe"
+  let AsmString = opcode#" $vdata, "#AddrAsm#", $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$a16$tfe$lwe"
                     #!if(BaseOpcode.HasD16, "$d16", "");
 }
 
@@ -331,9 +331,9 @@ class MIMG_Store_gfx10<int op, string opcode,
   : MIMG_gfx10<op, (outs), dns> {
   let InOperandList = !con((ins DataRC:$vdata, AddrRC:$vaddr0, SReg_256:$srsrc,
                                 DMask:$dmask, Dim:$dim, UNorm:$unorm, DLC:$dlc,
-                                GLC:$glc, SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe),
+                                GLC:$glc, SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe),
                            !if(BaseOpcode.HasD16, (ins D16:$d16), (ins)));
-  let AsmString = opcode#" $vdata, $vaddr0, $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$tfe$lwe"
+  let AsmString = opcode#" $vdata, $vaddr0, $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$a16$tfe$lwe"
                     #!if(BaseOpcode.HasD16, "$d16", "");
 }
 
@@ -345,9 +345,9 @@ class MIMG_Store_nsa_gfx10<int op, string opcode,
                            AddrIns,
                            (ins SReg_256:$srsrc, DMask:$dmask,
                                 Dim:$dim, UNorm:$unorm, DLC:$dlc, GLC:$glc,
-                                SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe),
+                                SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe),
                            !if(BaseOpcode.HasD16, (ins D16:$d16), (ins)));
-  let AsmString = opcode#" $vdata, "#AddrAsm#", $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$tfe$lwe"
+  let AsmString = opcode#" $vdata, "#AddrAsm#", $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$a16$tfe$lwe"
                     #!if(BaseOpcode.HasD16, "$d16", "");
 }
 
@@ -436,8 +436,8 @@ class MIMG_Atomic_gfx10<mimg op, string opcode,
 
   let InOperandList = (ins DataRC:$vdata, AddrRC:$vaddr0, SReg_256:$srsrc,
                            DMask:$dmask, Dim:$dim, UNorm:$unorm, DLC:$dlc,
-                           GLC:$glc, SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe);
-  let AsmString = opcode#" $vdst, $vaddr0, $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$tfe$lwe";
+                           GLC:$glc, SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe);
+  let AsmString = opcode#" $vdst, $vaddr0, $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$a16$tfe$lwe";
 }
 
 class MIMG_Atomic_nsa_gfx10<mimg op, string opcode,
@@ -452,8 +452,8 @@ class MIMG_Atomic_nsa_gfx10<mimg op, string opcode,
                            AddrIns,
                            (ins SReg_256:$srsrc, DMask:$dmask,
                                 Dim:$dim, UNorm:$unorm, DLC:$dlc, GLC:$glc,
-                                SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe));
-  let AsmString = opcode#" $vdata, "#AddrAsm#", $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$tfe$lwe";
+                                SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe));
+  let AsmString = opcode#" $vdata, "#AddrAsm#", $srsrc$dmask$dim$unorm$dlc$glc$slc$r128$a16$tfe$lwe";
 }
 
 multiclass MIMG_Atomic_Addr_Helper_m <mimg op, string asm,
@@ -522,10 +522,10 @@ class MIMG_Sampler_gfx10<int op, string opcode,
   : MIMG_gfx10<op, (outs DataRC:$vdata), dns> {
   let InOperandList = !con((ins AddrRC:$vaddr0, SReg_256:$srsrc, SReg_128:$ssamp,
                                 DMask:$dmask, Dim:$dim, UNorm:$unorm, DLC:$dlc,
-                                GLC:$glc, SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe),
+                                GLC:$glc, SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe),
                            !if(BaseOpcode.HasD16, (ins D16:$d16), (ins)));
   let AsmString = opcode#" $vdata, $vaddr0, $srsrc, $ssamp$dmask$dim$unorm"
-                    #"$dlc$glc$slc$r128$tfe$lwe"
+                    #"$dlc$glc$slc$r128$a16$tfe$lwe"
                     #!if(BaseOpcode.HasD16, "$d16", "");
 }
 
@@ -536,10 +536,10 @@ class MIMG_Sampler_nsa_gfx10<int op, string opcode,
   let InOperandList = !con(AddrIns,
                            (ins SReg_256:$srsrc, SReg_128:$ssamp, DMask:$dmask,
                                 Dim:$dim, UNorm:$unorm, DLC:$dlc, GLC:$glc,
-                                SLC:$slc, R128A16:$r128, TFE:$tfe, LWE:$lwe),
+                                SLC:$slc, R128A16:$r128, GFX10A16:$a16, TFE:$tfe, LWE:$lwe),
                            !if(BaseOpcode.HasD16, (ins D16:$d16), (ins)));
   let AsmString = opcode#" $vdata, "#AddrAsm#", $srsrc, $ssamp$dmask$dim$unorm"
-                    #"$dlc$glc$slc$r128$tfe$lwe"
+                    #"$dlc$glc$slc$r128$a16$tfe$lwe"
                     #!if(BaseOpcode.HasD16, "$d16", "");
 }
 
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 475a2b8..e60ad75 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -5401,7 +5401,7 @@ SDValue SITargetLowering::lowerImage(SDValue Op,
   const MVT VAddrScalarVT = VAddrVT.getScalarType();
   if (((VAddrScalarVT == MVT::f16) || (VAddrScalarVT == MVT::i16))) {
     // Illegal to use a16 images
-    if (!ST->hasFeature(AMDGPU::FeatureR128A16))
+    if (!ST->hasFeature(AMDGPU::FeatureR128A16) && !ST->hasFeature(AMDGPU::FeatureGFX10A16))
       return Op;
 
     IsA16 = true;
@@ -5546,10 +5546,12 @@ SDValue SITargetLowering::lowerImage(SDValue Op,
     Ops.push_back(DLC);
   Ops.push_back(GLC);
   Ops.push_back(SLC);
-  Ops.push_back(IsA16 &&  // a16 or r128
+  Ops.push_back(IsA16 &&  // r128, a16 for gfx9
                 ST->hasFeature(AMDGPU::FeatureR128A16) ? True : False);
-  Ops.push_back(TFE); // tfe
-  Ops.push_back(LWE); // lwe
+  if (IsGFX10)
+    Ops.push_back(IsA16 ? True : False);
+  Ops.push_back(TFE);
+  Ops.push_back(LWE);
   if (!IsGFX10)
     Ops.push_back(DimInfo->DA ? True : False);
   if (BaseOpcode->HasD16)
diff --git a/llvm/lib/Target/AMDGPU/SIInstrFormats.td b/llvm/lib/Target/AMDGPU/SIInstrFormats.td
index 4dcbe92..a4d1178 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrFormats.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrFormats.td
@@ -303,7 +303,7 @@ class MIMGe_gfx10 <bits<8> op> : MIMGe {
   bits<3> dim;
   bits<2> nsa;
   bits<1> dlc;
-  bits<1> a16 = 0; // TODO: this should be an operand
+  bits<1> a16;
 
   let Inst{0} = op{7};
   let Inst{2-1} = nsa;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index a9800bb..73690ae 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -3742,11 +3742,34 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
         return false;
       }
 
+      bool IsA16 = false;
+      if (ST.hasR128A16()) {
+        const MachineOperand *R128A16 = getNamedOperand(MI, AMDGPU::OpName::r128);
+        IsA16 = R128A16->getImm() != 0;
+      } else if (ST.hasGFX10A16()) {
+        const MachineOperand *A16 = getNamedOperand(MI, AMDGPU::OpName::a16);
+        IsA16 = A16->getImm() != 0;
+      }
+
+      bool PackDerivatives = IsA16; // Either A16 or G16
       bool IsNSA = SRsrcIdx - VAddr0Idx > 1;
-      unsigned AddrWords = BaseOpcode->NumExtraArgs +
-                           (BaseOpcode->Gradients ? Dim->NumGradients : 0) +
-                           (BaseOpcode->Coordinates ? Dim->NumCoords : 0) +
-                           (BaseOpcode->LodOrClampOrMip ? 1 : 0);
+
+      unsigned AddrWords = BaseOpcode->NumExtraArgs;
+      unsigned AddrComponents = (BaseOpcode->Coordinates ? Dim->NumCoords : 0) +
+                                (BaseOpcode->LodOrClampOrMip ? 1 : 0);
+      if (IsA16)
+        AddrWords += (AddrComponents + 1) / 2;
+      else
+        AddrWords += AddrComponents;
+
+      if (BaseOpcode->Gradients) {
+        if (PackDerivatives)
+          // There are two gradients per coordinate, we pack them separately.
+          // For the 3d case, we get (dy/du, dx/du) (-, dz/du) (dy/dv, dx/dv) (-, dz/dv)
+          AddrWords += (Dim->NumGradients / 2 + 1) / 2 * 2;
+        else
+          AddrWords += Dim->NumGradients;
+      }
 
       unsigned VAddrWords;
       if (IsNSA) {
@@ -3758,11 +3781,10 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
           AddrWords = 16;
         else if (AddrWords > 4)
           AddrWords = 8;
-        else if (AddrWords == 3 && VAddrWords == 4) {
-          // CodeGen uses the V4 variant of instructions for three addresses,
-          // because the selection DAG does not support non-power-of-two types.
+        else if (AddrWords == 4)
           AddrWords = 4;
-        }
+        else if (AddrWords == 3)
+          AddrWords = 3;
       }
 
       if (VAddrWords != AddrWords) {
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index bf6f029..559ede8 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -1090,6 +1090,7 @@ def SWZ : NamedOperandBit<"SWZ", NamedMatchClass<"SWZ">>;
 def UNorm : NamedOperandBit<"UNorm", NamedMatchClass<"UNorm">>;
 def DA : NamedOperandBit<"DA", NamedMatchClass<"DA">>;
 def R128A16 : NamedOperandBit<"R128A16", NamedMatchClass<"R128A16">>;
+def GFX10A16 : NamedOperandBit<"GFX10A16", NamedMatchClass<"GFX10A16">>;
 def D16 : NamedOperandBit<"D16", NamedMatchClass<"D16">>;
 def LWE : NamedOperandBit<"LWE", NamedMatchClass<"LWE">>;
 def exp_compr : NamedOperandBit<"ExpCompr", NamedMatchClass<"ExpCompr">>;
diff --git a/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp b/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
index 806f8af..dd94a29 100644
--- a/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
+++ b/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
@@ -686,7 +686,8 @@ bool SILoadStoreOptimizer::dmasksCanBeCombined(const CombineInfo &CI,
   // Check other optional immediate operands for equality.
   unsigned OperandsToMatch[] = {AMDGPU::OpName::glc, AMDGPU::OpName::slc,
                                 AMDGPU::OpName::d16, AMDGPU::OpName::unorm,
-                                AMDGPU::OpName::da,  AMDGPU::OpName::r128};
+                                AMDGPU::OpName::da,  AMDGPU::OpName::r128,
+                                AMDGPU::OpName::a16};
 
   for (auto op : OperandsToMatch) {
     int Idx = AMDGPU::getNamedOperandIdx(CI.I->getOpcode(), op);
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index 1188671..a5ed441 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -927,7 +927,11 @@ bool hasSRAMECC(const MCSubtargetInfo &STI) {
 }
 
 bool hasMIMG_R128(const MCSubtargetInfo &STI) {
-  return STI.getFeatureBits()[AMDGPU::FeatureMIMG_R128];
+  return STI.getFeatureBits()[AMDGPU::FeatureMIMG_R128] && !STI.getFeatureBits()[AMDGPU::FeatureR128A16];
+}
+
+bool hasGFX10A16(const MCSubtargetInfo &STI) {
+  return STI.getFeatureBits()[AMDGPU::FeatureGFX10A16];
 }
 
 bool hasPackedD16(const MCSubtargetInfo &STI) {
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index da2f4dc..4cfb2ad 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -551,6 +551,7 @@ inline bool isKernel(CallingConv::ID CC) {
 bool hasXNACK(const MCSubtargetInfo &STI);
 bool hasSRAMECC(const MCSubtargetInfo &STI);
 bool hasMIMG_R128(const MCSubtargetInfo &STI);
+bool hasGFX10A16(const MCSubtargetInfo &STI);
 bool hasPackedD16(const MCSubtargetInfo &STI);
 
 bool isSI(const MCSubtargetInfo &STI);