AMDGPU/SI: Add support for 32-bit immediate SMRD offsets on SI

[arsenm]

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

1581

Typo: subtaget

1587–1589

I think you need to be careful here if the immediate is == 255. It is still is uint<8>, but the encoding value will end up reading the next 4 bytes of instruction instead of the right value.

test/CodeGen/AMDGPU/smrd.ll

19

These should probably be split into SI and CI lines. I think on SI this will end up using a 32-bit encoding and will have to use on a 64-bit one on CI for 255

test/MC/AMDGPU/smrd.s

11

Should probably include a test with 0xff too

[arsenm]

Title / commit message should also say CI

[• tstellarAMD]

Fixed typo and improve test for offset equal to 0xff.

[• tstellarAMD]

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

1587–1589

Are you concerned that the value 0xff will be read as SRC_LITERAL rather than the actual offset? Because there is no way to specify SRC_LITERAL using the assembler.

test/CodeGen/AMDGPU/smrd.ll

19

I double -checked and 32-bit immediates are only supported on CI. I updated this test to make sure it is using the 32-bit encoding of SMRD.

[arsenm]

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

1587–1589

Yes. My understanding was you can't use 255 because that is the same as SRC_LITERAL and both are set in the same immediate field. The meaning of 255 would then change meaning depending on SI/CI.

test/CodeGen/AMDGPU/smrd.ll

19

Because 255 is SQ_LITERAL on CI I think it will be 64-bit there

[• tstellarAMD]

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

1587–1589

The meaning of 255 is controled by the IMM bit, so there shouldn't be an issue.
The encoding looks like this:

8-bit offset (0xff)
offset = 0xff
imm = 1

32-bit offset
offset = 0xff
imm = 0

[arsenm]

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

1587–1589

OK, I didn't notice that bit. I think the assembler should probably just be smart about it and turn it on only if it sees an immediate that won't fit in 8-bits

[• tstellarAMD]

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

1587–1589

The assembler already does this.

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
                          SDValue &TFE) const;
   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
                          SDValue &Offset, SDValue &GLC) const;
   bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue &Offset,
                         bool &Imm) const;
   bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue &Offset,
                   bool &Imm) const;
   bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
+  bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
   bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
   bool SelectSMRDBufferImm(SDValue Addr, SDValue &Offset) const;
+  bool SelectSMRDBufferImm32(SDValue Addr, SDValue &Offset) const;
   bool SelectSMRDBufferSgpr(SDValue Addr, SDValue &Offset) const;
   SDNode *SelectAddrSpaceCast(SDNode *N);
   bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
   bool SelectVOP3NoMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
   bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
                        SDValue &Clamp, SDValue &Omod) const;
   bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
                          SDValue &Clamp, SDValue &Omod) const;
int64_t EncodedOffset = Gen < AMDGPUSubtarget::VOLCANIC_ISLANDS ?
       ByteOffset >> 2 : ByteOffset;
 
   if (isLegalSMRDImmOffset(Subtarget, EncodedOffset)) {
     Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32);
     Imm = true;
     return true;
   }
 
-  if (isUInt<32>(ByteOffset)) {
+  if (!isUInt<32>(EncodedOffset) || !isUInt<32>(ByteOffset))
+    return false;
+
+  if (Gen == AMDGPUSubtarget::SEA_ISLANDS && isUInt<32>(EncodedOffset)) {
+    // 32-bit Immediates are supported on Sea Islands.
+    Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32);
+  } else {
     SDValue C32Bit = CurDAG->getTargetConstant(ByteOffset, SL, MVT::i32);
     Offset = SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32,
                                             C32Bit), 0);
-    Imm = false;
-    return true;
-  }
-  return false;
-}
+  }
 
 bool AMDGPUDAGToDAGISel::SelectSMRD(SDValue Addr, SDValue &SBase,
                                      SDValue &Offset, bool &Imm) const {
 
   SDLoc SL(Addr);
   if (CurDAG->isBaseWithConstantOffset(Addr)) {
     SDValue N0 = Addr.getOperand(0);
     SDValue N1 = Addr.getOperand(1);

 }
 
 bool AMDGPUDAGToDAGISel::SelectSMRDImm(SDValue Addr, SDValue &SBase,
                                        SDValue &Offset) const {
   bool Imm;
   return SelectSMRD(Addr, SBase, Offset, Imm) && Imm;
 }
 
+bool AMDGPUDAGToDAGISel::SelectSMRDImm32(SDValue Addr, SDValue &SBase,
+                                         SDValue &Offset) const {
+
+  if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS)
+    return false;
+
+  bool Imm;
+  if (!SelectSMRD(Addr, SBase, Offset, Imm))
+    return false;
+
+  return !Imm && isa<ConstantSDNode>(Offset);
+}
+
 bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase,
                                         SDValue &Offset) const {
   bool Imm;
-  return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm;
+  return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm &&
+         !isa<ConstantSDNode>(Offset);
 }
 
 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr,
                                              SDValue &Offset) const {
   bool Imm;
   return SelectSMRDOffset(Addr, Offset, Imm) && Imm;
 }
 
+bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm32(SDValue Addr,
+                                               SDValue &Offset) const {
+  if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS)
+    return false;
+
+  bool Imm;
+  if (!SelectSMRDOffset(Addr, Offset, Imm))
+    return false;
+
+  return !Imm && isa<ConstantSDNode>(Offset);
+}
+
 bool AMDGPUDAGToDAGISel::SelectSMRDBufferSgpr(SDValue Addr,
                                               SDValue &Offset) const {
   bool Imm;
-  return SelectSMRDOffset(Addr, Offset, Imm) && !Imm;
+  return SelectSMRDOffset(Addr, Offset, Imm) && !Imm &&
+         !isa<ConstantSDNode>(Offset);
 }
 
 // FIXME: This is incorrect and only enough to be able to compile.
 SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
   AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
   SDLoc DL(N);
 
   assert(Subtarget->hasFlatAddressSpace() &&

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

static std::unique_ptr<AMDGPUOperand> CreateExpr(const class MCExpr *Expr, SMLoc S) {
     Op->EndLoc = S;
     return Op;
   }
 
   bool isDSOffset() const;
   bool isDSOffset01() const;
   bool isSWaitCnt() const;
   bool isMubufOffset() const;
+  bool isSMRDOffset() const;
+  bool isSMRDLiteralOffset() const;
 };
 
 class AMDGPUAsmParser : public MCTargetAsmParser {
   MCSubtargetInfo &STI;
   const MCInstrInfo &MII;
   MCAsmParser &Parser;
 
   unsigned ForcedEncodingSize;

 }
 
 AMDGPUAsmParser::OperandMatchResultTy
 AMDGPUAsmParser::parseR128(OperandVector &Operands) {
   return parseNamedBit("r128", Operands);
 }
 
 //===----------------------------------------------------------------------===//
+// smrd
+//===----------------------------------------------------------------------===//
+
+bool AMDGPUOperand::isSMRDOffset() const {
+
+  // FIXME: Support 20-bit offsets on VI.  We need to to pass subtarget
+  // information here.
+  return isImm() && isUInt<8>(getImm());
+}
+
+bool AMDGPUOperand::isSMRDLiteralOffset() const {
+  // 32-bit literals are only supported on CI and we only want to use them
+  // when the offset is > 8-bits.
+  return isImm() && !isUInt<8>(getImm()) && isUInt<32>(getImm());
+}
+
+//===----------------------------------------------------------------------===//
 // vop3
 //===----------------------------------------------------------------------===//
 
 static bool ConvertOmodMul(int64_t &Mul) {
   if (Mul != 1 && Mul != 2 && Mul != 4)
     return false;
 
   Mul >>= 1;

lib/Target/AMDGPU/SIInstrFormats.td

class SMRD_IMMe_ci <bits<5> op> : Enc64 {
   bits<7> sdst;
   bits<7> sbase;
   bits<32> offset;
 
   let Inst{7-0}   = 0xff;
   let Inst{8}     = 0;
   let Inst{14-9}  = sbase{6-1};
   let Inst{21-15} = sdst;
-  let Inst{26-22} = 0x18; //encoding
+  let Inst{26-22} = op;
+  let Inst{31-27} = 0x18; //encoding
   let Inst{63-32} = offset;
 }
 
 let SchedRW = [WriteSALU] in {
 class SOP1 <dag outs, dag ins, string asm, list<dag> pattern> :
     InstSI<outs, ins, asm, pattern> {
   let mayLoad = 0;
   let mayStore = 0;

lib/Target/AMDGPU/SIInstrInfo.td

 //===-- SIInstrInfo.td - SI Instruction Infos -------------*- tablegen -*--===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 def isCI : Predicate<"Subtarget->getGeneration() "
                       ">= AMDGPUSubtarget::SEA_ISLANDS">;
+def isCIOnly : Predicate<"Subtarget->getGeneration() =="
+                         "AMDGPUSubtarget::SEA_ISLANDS">,
+  AssemblerPredicate <"FeatureSeaIslands">;
 def isVI : Predicate <
   "Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS">,
   AssemblerPredicate<"FeatureGCN3Encoding">;
 
 def DisableInst : Predicate <"false">, AssemblerPredicate<"FeatureDisable">;
 
 class vop {
   field bits<9> SI3;

 
 def ClampMatchClass : AsmOperandClass {
   let Name = "Clamp";
   let PredicateMethod = "isImm";
   let ParserMethod = "parseVOP3OptionalOps";
   let RenderMethod = "addImmOperands";
 }
 
+class SMRDOffsetBaseMatchClass <string predicate> : AsmOperandClass {
+  let Name = "SMRDOffset"#predicate;
+  let PredicateMethod = predicate;
+  let RenderMethod = "addImmOperands";
+}
+
+def SMRDOffsetMatchClass : SMRDOffsetBaseMatchClass <"isSMRDOffset">;
+def SMRDLiteralOffsetMatchClass : SMRDOffsetBaseMatchClass <
+  "isSMRDLiteralOffset"
+>;
+
 let OperandType = "OPERAND_IMMEDIATE" in {
 
 def offen : Operand<i1> {
   let PrintMethod = "printOffen";
 }
 def idxen : Operand<i1> {
   let PrintMethod = "printIdxen";
 }
def omod : Operand <i32> {
   let ParserMatchClass = OModMatchClass;
 }
 
 def ClampMod : Operand <i1> {
   let PrintMethod = "printClampSI";
   let ParserMatchClass = ClampMatchClass;
 }
 
+def smrd_offset : Operand <i32> {
+  let PrintMethod = "printU32ImmOperand";
+  let ParserMatchClass = SMRDOffsetMatchClass;
+}
+
+def smrd_literal_offset : Operand <i32> {
+  let PrintMethod = "printU32ImmOperand";
+  let ParserMatchClass = SMRDLiteralOffsetMatchClass;
+}
+
 } // End OperandType = "OPERAND_IMMEDIATE"
 
 def VOPDstS64 : VOPDstOperand <SReg_64>;
 
 //===----------------------------------------------------------------------===//
 // Complex patterns
 //===----------------------------------------------------------------------===//
 
 def DS1Addr1Offset : ComplexPattern<i32, 2, "SelectDS1Addr1Offset">;
 def DS64Bit4ByteAligned : ComplexPattern<i32, 3, "SelectDS64Bit4ByteAligned">;
 
 def MUBUFAddr32 : ComplexPattern<i64, 9, "SelectMUBUFAddr32">;
 def MUBUFAddr64 : ComplexPattern<i64, 7, "SelectMUBUFAddr64">;
 def MUBUFAddr64Atomic : ComplexPattern<i64, 5, "SelectMUBUFAddr64">;
 def MUBUFScratch : ComplexPattern<i64, 4, "SelectMUBUFScratch">;
 def MUBUFOffset : ComplexPattern<i64, 6, "SelectMUBUFOffset">;
 def MUBUFOffsetAtomic : ComplexPattern<i64, 4, "SelectMUBUFOffset">;
 
 def SMRDImm   : ComplexPattern<i64, 2, "SelectSMRDImm">;
+def SMRDImm32 : ComplexPattern<i64, 2, "SelectSMRDImm32">;
 def SMRDSgpr  : ComplexPattern<i64, 2, "SelectSMRDSgpr">;
 def SMRDBufferImm   : ComplexPattern<i32, 1, "SelectSMRDBufferImm">;
+def SMRDBufferImm32 : ComplexPattern<i32, 1, "SelectSMRDBufferImm32">;
 def SMRDBufferSgpr  : ComplexPattern<i32, 1, "SelectSMRDBufferSgpr">;
 
-
 def VOP3Mods0 : ComplexPattern<untyped, 4, "SelectVOP3Mods0">;
 def VOP3NoMods0 : ComplexPattern<untyped, 4, "SelectVOP3NoMods0">;
 def VOP3Mods0Clamp : ComplexPattern<untyped, 3, "SelectVOP3Mods0Clamp">;
 def VOP3Mods0Clamp0OMod : ComplexPattern<untyped, 4, "SelectVOP3Mods0Clamp0OMod">;
 def VOP3Mods  : ComplexPattern<untyped, 2, "SelectVOP3Mods">;
 def VOP3NoMods : ComplexPattern<untyped, 2, "SelectVOP3NoMods">;
 
 //===----------------------------------------------------------------------===//
let glc = 0 in {
     def _vi : SMRD_Real_vi <{0, 0, 0, op}, opName, imm, outs, ins, asm>;
   }
 }
 
 multiclass SMRD_Helper <bits<5> op, string opName, RegisterClass baseClass,
                         RegisterClass dstClass> {
   defm _IMM : SMRD_m <
     op, opName#"_IMM", 1, (outs dstClass:$dst),
-    (ins baseClass:$sbase, u32imm:$offset),
+    (ins baseClass:$sbase, smrd_offset:$offset),
     opName#" $dst, $sbase, $offset", []
   >;
 
   def _IMM_ci : SMRD <
-    (outs dstClass:$dst), (ins baseClass:$sbase, u32imm:$offset),
+    (outs dstClass:$dst), (ins baseClass:$sbase, smrd_literal_offset:$offset),
     opName#" $dst, $sbase, $offset", []>, SMRD_IMMe_ci <op> {
-    let AssemblerPredicates = [isCI];
+    let AssemblerPredicates = [isCIOnly];
   }
 
   defm _SGPR : SMRD_m <
     op, opName#"_SGPR", 0, (outs dstClass:$dst),
     (ins baseClass:$sbase, SReg_32:$soff),
     opName#" $dst, $sbase, $soff", []
   >;
 }

lib/Target/AMDGPU/SIInstructions.td

def : Pat <
   (EXP imm:$en, imm:$tgt, imm:$compr, imm:$done, imm:$vm,
        $src0, $src1, $src2, $src3)
 >;
 
 //===----------------------------------------------------------------------===//
 // SMRD Patterns
 //===----------------------------------------------------------------------===//
 
-multiclass SMRD_Pattern <SMRD Instr_IMM, SMRD Instr_SGPR, ValueType vt> {
+multiclass SMRD_Pattern <string Instr, ValueType vt> {
 
   // 1. IMM offset
   def : Pat <
     (constant_load (SMRDImm i64:$sbase, i32:$offset)),
-    (vt (Instr_IMM $sbase, $offset))
+    (vt (!cast<SMRD>(Instr#"_IMM") $sbase, $offset))
   >;
 
   // 2. SGPR offset
   def : Pat <
     (constant_load (SMRDSgpr i64:$sbase, i32:$offset)),
-    (vt (Instr_SGPR $sbase, $offset))
+    (vt (!cast<SMRD>(Instr#"_SGPR") $sbase, $offset))
   >;
+
+  def : Pat <
+    (constant_load (SMRDImm32 i64:$sbase, i32:$offset)),
+    (vt (!cast<SMRD>(Instr#"_IMM_ci") $sbase, $offset))
+  > {
+    let Predicates = [isCIOnly];
+  }
 }
 
-defm : SMRD_Pattern <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, i32>;
-defm : SMRD_Pattern <S_LOAD_DWORDX2_IMM, S_LOAD_DWORDX2_SGPR, v2i32>;
-defm : SMRD_Pattern <S_LOAD_DWORDX4_IMM, S_LOAD_DWORDX4_SGPR, v4i32>;
-defm : SMRD_Pattern <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v32i8>;
-defm : SMRD_Pattern <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v8i32>;
-defm : SMRD_Pattern <S_LOAD_DWORDX16_IMM, S_LOAD_DWORDX16_SGPR, v16i32>;
+defm : SMRD_Pattern <"S_LOAD_DWORD", i32>;
+defm : SMRD_Pattern <"S_LOAD_DWORDX2", v2i32>;
+defm : SMRD_Pattern <"S_LOAD_DWORDX4", v4i32>;
+defm : SMRD_Pattern <"S_LOAD_DWORDX8", v32i8>;
+defm : SMRD_Pattern <"S_LOAD_DWORDX8", v8i32>;
+defm : SMRD_Pattern <"S_LOAD_DWORDX16", v16i32>;
 
 // 1. Offset as an immediate
 def : Pat <
   (SIload_constant v4i32:$sbase, (SMRDBufferImm i32:$offset)),
   (S_BUFFER_LOAD_DWORD_IMM $sbase, $offset)
 >;
 
 // 2. Offset loaded in an 32bit SGPR
 def : Pat <
   (SIload_constant v4i32:$sbase, (SMRDBufferSgpr i32:$offset)),
   (S_BUFFER_LOAD_DWORD_SGPR $sbase, $offset)
 >;
 
+let Predicates = [isCI] in {
+
+def : Pat <
+  (SIload_constant v4i32:$sbase, (SMRDBufferImm32 i32:$offset)),
+  (S_BUFFER_LOAD_DWORD_IMM_ci $sbase, $offset)
+>;
+
+} // End Predicates = [isCI]
+
 //===----------------------------------------------------------------------===//
 // SOP1 Patterns
 //===----------------------------------------------------------------------===//
 
 def : Pat <
   (i64 (ctpop i64:$src)),
     (i64 (REG_SEQUENCE SReg_64,
      (S_BCNT1_I32_B64 $src), sub0,

test/CodeGen/AMDGPU/smrd.ll

-; RUN: llc < %s -march=amdgcn -mcpu=SI -show-mc-encoding -verify-machineinstrs | FileCheck --check-prefix=SI --check-prefix=GCN %s
-; RUN: llc < %s -march=amdgcn -mcpu=tonga -show-mc-encoding -verify-machineinstrs | FileCheck --check-prefix=VI --check-prefix=GCN %s
+; RUN: llc < %s -march=amdgcn -mcpu=SI -show-mc-encoding -verify-machineinstrs | FileCheck --check-prefix=SI --check-prefix=GCN --check-prefix=SIVI %s
+; RUN: llc < %s -march=amdgcn -mcpu=bonaire -show-mc-encoding -verify-machineinstrs | FileCheck --check-prefix=CI --check-prefix=GCN  %s
+; RUN: llc < %s -march=amdgcn -mcpu=tonga -show-mc-encoding -verify-machineinstrs | FileCheck --check-prefix=VI --check-prefix=GCN --check-prefix=SIVI %s
 
 ; SMRD load with an immediate offset.
 ; GCN-LABEL: {{^}}smrd0:
-; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x1 ; encoding: [0x01
+; SICI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x1 ; encoding: [0x01
 ; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4
 define void @smrd0(i32 addrspace(1)* %out, i32 addrspace(2)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(2)* %ptr, i64 1
   %1 = load i32, i32 addrspace(2)* %0
   store i32 %1, i32 addrspace(1)* %out
   ret void
 }
 
 ; SMRD load with the largest possible immediate offset.
 ; GCN-LABEL: {{^}}smrd1:
-; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff
+; SICI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff,0x{{[0-9]+[137]}}
 ; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
 define void @smrd1(i32 addrspace(1)* %out, i32 addrspace(2)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(2)* %ptr, i64 255
   %1 = load i32, i32 addrspace(2)* %0
   store i32 %1, i32 addrspace(1)* %out
   ret void
 }
 
 ; SMRD load with an offset greater than the largest possible immediate.
 ; GCN-LABEL: {{^}}smrd2:
 ; SI: s_movk_i32 s[[OFFSET:[0-9]]], 0x400
 ; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
+; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
 ; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
 ; GCN: s_endpgm
 define void @smrd2(i32 addrspace(1)* %out, i32 addrspace(2)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(2)* %ptr, i64 256
   %1 = load i32, i32 addrspace(2)* %0
   store i32 %1, i32 addrspace(1)* %out
   ret void
entry:
   store i32 %1, i32 addrspace(1)* %out
   ret void
 }
 
 ; SMRD load with the largest possible immediate offset on VI
 ; GCN-LABEL: {{^}}smrd4:
 ; SI: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
 ; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
 ; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
 define void @smrd4(i32 addrspace(1)* %out, i32 addrspace(2)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(2)* %ptr, i64 262143
   %1 = load i32, i32 addrspace(2)* %0
   store i32 %1, i32 addrspace(1)* %out
   ret void
 }
 
 ; SMRD load with an offset greater than the largest possible immediate on VI
 ; GCN-LABEL: {{^}}smrd5:
-; GCN: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
-; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
-; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; SIVI: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
+; SIVI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
 ; GCN: s_endpgm
 define void @smrd5(i32 addrspace(1)* %out, i32 addrspace(2)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(2)* %ptr, i64 262144
   %1 = load i32, i32 addrspace(2)* %0
   store i32 %1, i32 addrspace(1)* %out
   ret void
 }
 
 ; SMRD load using the load.const intrinsic with an immediate offset
 ; GCN-LABEL: {{^}}smrd_load_const0:
-; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4 ; encoding: [0x04
+; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4 ; encoding: [0x04
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x10
 define void @smrd_load_const0(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) #0 {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 16)
   call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
   ret void
 }
 
 ; SMRD load using the load.const intrinsic with the largest possible immediate
 ; offset.
 ; GCN-LABEL: {{^}}smrd_load_const1:
-; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff
+; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
 define void @smrd_load_const1(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) #0 {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1020)
   call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
   ret void
 }
 ; SMRD load using the load.const intrinsic with an offset greater than the
 ; largets possible immediate.
 ; immediate offset.
 ; GCN-LABEL: {{^}}smrd_load_const2:
 ; SI: s_movk_i32 s[[OFFSET:[0-9]]], 0x400
 ; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
+; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
 define void @smrd_load_const2(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) #0 {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1024)
   call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
   ret void
 }
 
 ; SMRD load with the largest possible immediate offset on VI
 ; GCN-LABEL: {{^}}smrd_load_const3:
 ; SI: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
 ; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
 define void @smrd_load_const3(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) #0 {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1048572)
   call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
   ret void
 }
 
 ; SMRD load with an offset greater than the largest possible immediate on VI
 ; GCN-LABEL: {{^}}smrd_load_const4:
-; GCN: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
-; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
-; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; SIVI: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
+; SIVI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
 ; GCN: s_endpgm
 define void @smrd_load_const4(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) #0 {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1048576)
   call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
   ret void

test/MC/AMDGPU/smrd.s

-// RUN: llvm-mc -arch=amdgcn -show-encoding %s | FileCheck %s
-// RUN: llvm-mc -arch=amdgcn -mcpu=SI -show-encoding %s | FileCheck %s
+// RUN: not llvm-mc -arch=amdgcn -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SI  %s
+// RUN: not llvm-mc -arch=amdgcn -mcpu=SI -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SI %s
+// RUN: llvm-mc -arch=amdgcn -mcpu=bonaire -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=CI %s
+
+// RUN: not llvm-mc -arch=amdgcn %s 2>&1 | FileCheck %s --check-prefix=NOSI
+// RUN: not llvm-mc -arch=amdgcn -mcpu=SI  %s 2>&1 | FileCheck %s --check-prefix=NOSI
+//===----------------------------------------------------------------------===//
+// Offset Handling
+//===----------------------------------------------------------------------===//
+
+s_load_dword s1, s[2:3], 0xfc
+// GCN: s_load_dword s1, s[2:3], 0xfc ; encoding: [0xfc,0x83,0x00,0xc0]
+
+s_load_dword s1, s[2:3], 0xff
+// GCN: s_load_dword s1, s[2:3], 0xff ; encoding: [0xff,0x83,0x00,0xc0]
+
+s_load_dword s1, s[2:3], 0x100
+// NOSI: error: instruction not supported on this GPU
+// CI: s_load_dword s1, s[2:3], 0x100 ; encoding: [0xff,0x82,0x00,0xc0,0x00,0x01,0x00,0x00]
+
+//===----------------------------------------------------------------------===//
+// Instructions
+//===----------------------------------------------------------------------===//
 
 s_load_dword s1, s[2:3], 1
-// CHECK: s_load_dword s1, s[2:3], 0x1 ; encoding: [0x01,0x83,0x00,0xc0]
+// GCN: s_load_dword s1, s[2:3], 0x1 ; encoding: [0x01,0x83,0x00,0xc0]
 
 s_load_dword s1, s[2:3], s4
-// CHECK: s_load_dword s1, s[2:3], s4 ; encoding: [0x04,0x82,0x00,0xc0]
+// GCN: s_load_dword s1, s[2:3], s4 ; encoding: [0x04,0x82,0x00,0xc0]
 
 s_load_dwordx2 s[2:3], s[2:3], 1
-// CHECK: s_load_dwordx2 s[2:3], s[2:3], 0x1 ; encoding: [0x01,0x03,0x41,0xc0]
+// GCN: s_load_dwordx2 s[2:3], s[2:3], 0x1 ; encoding: [0x01,0x03,0x41,0xc0]
 
 s_load_dwordx2 s[2:3], s[2:3], s4
-// CHECK: s_load_dwordx2 s[2:3], s[2:3], s4 ; encoding: [0x04,0x02,0x41,0xc0]
+// GCN: s_load_dwordx2 s[2:3], s[2:3], s4 ; encoding: [0x04,0x02,0x41,0xc0]
 
 s_load_dwordx4 s[4:7], s[2:3], 1
-// CHECK: s_load_dwordx4 s[4:7], s[2:3], 0x1 ; encoding: [0x01,0x03,0x82,0xc0]
+// GCN: s_load_dwordx4 s[4:7], s[2:3], 0x1 ; encoding: [0x01,0x03,0x82,0xc0]
 
 s_load_dwordx4 s[4:7], s[2:3], s4
-// CHECK: s_load_dwordx4 s[4:7], s[2:3], s4 ; encoding: [0x04,0x02,0x82,0xc0]
+// GCN: s_load_dwordx4 s[4:7], s[2:3], s4 ; encoding: [0x04,0x02,0x82,0xc0]
 
 s_load_dwordx8 s[8:15], s[2:3], 1
-// CHECK: s_load_dwordx8 s[8:15], s[2:3], 0x1 ; encoding: [0x01,0x03,0xc4,0xc0]
+// GCN: s_load_dwordx8 s[8:15], s[2:3], 0x1 ; encoding: [0x01,0x03,0xc4,0xc0]
 
 s_load_dwordx8 s[8:15], s[2:3], s4
-// CHECK: s_load_dwordx8 s[8:15], s[2:3], s4 ; encoding: [0x04,0x02,0xc4,0xc0]
+// GCN: s_load_dwordx8 s[8:15], s[2:3], s4 ; encoding: [0x04,0x02,0xc4,0xc0]
 
 s_load_dwordx16 s[16:31], s[2:3], 1
-// CHECK: s_load_dwordx16 s[16:31], s[2:3], 0x1 ; encoding: [0x01,0x03,0x08,0xc1]
+// GCN: s_load_dwordx16 s[16:31], s[2:3], 0x1 ; encoding: [0x01,0x03,0x08,0xc1]
 
 s_load_dwordx16 s[16:31], s[2:3], s4
-// CHECK: s_load_dwordx16 s[16:31], s[2:3], s4 ; encoding: [0x04,0x02,0x08,0xc1]
+// GCN: s_load_dwordx16 s[16:31], s[2:3], s4 ; encoding: [0x04,0x02,0x08,0xc1]