AMDGPU/SI: Use ComplexPatterns for SMRD addressing modes

[• tstellarAMD]

Summary

This allows us to consolidate several of the TableGen patterns.

[arsenm]

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1164–1167

I don't think this is accurate for VI. On VI, these instructions are called SMEM format instead of SMRD, so the function name isn't quite right. More importantly, I thought the 20-bit offset on VI was in bytes and dwords on SI/CI.

[• tstellarAMD]

Updated patch to clarify input to isLegalSMRDImmOffset()

[• tstellarAMD]

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1164–1167

The value passed to the function is the encoded value, so it is already in the correct units for the target. SMRD was renamed to SMEM, but the load instructions are still identical, and we call them SMRD in a few other places.

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1165–1166

I would add another sentence "Note that the offset has different units on VI" or something like that

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
                          SDValue &SLC) const;
   bool SelectMUBUFScratch(SDValue Addr, SDValue &RSrc, SDValue &VAddr,
                           SDValue &SOffset, SDValue &ImmOffset) const;
   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
                          SDValue &Offset, SDValue &GLC, SDValue &SLC,
                          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 SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
+  bool SelectSMRDBufferImm(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;
 

 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
                                            SDValue &Soffset, SDValue &Offset,
                                            SDValue &GLC) const {
   SDValue SLC, TFE;
 
   return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE);
 }
 
+///
+/// \param EncodedOffset This is the immediate value that will be encoded
+///        directly into the instruction.
+static bool isLegalSMRDImmOffset(const AMDGPUSubtarget *ST,
+                                 int64_t EncodedOffset) {
+  return ST->getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS ?
+     isUInt<8>(Offset) : isUInt<20>(Offset);
+}
+
+bool AMDGPUDAGToDAGISel::SelectSMRDOffset(SDValue ByteOffsetNode,
+                                          SDValue &Offset, bool &Imm) const {
+
+  // FIXME: Handle non-constant offsets.
+  ConstantSDNode *C = dyn_cast<ConstantSDNode>(ByteOffsetNode);
+  if (!C)
+    return false;
+
+  SDLoc SL(ByteOffsetNode);
+  AMDGPUSubtarget::Generation Gen = Subtarget->getGeneration();
+  int64_t ByteOffset = C->getSExtValue();
+  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)) {
+    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);
+
+    if (SelectSMRDOffset(N1, Offset, Imm)) {
+      SBase = N0;
+      return true;
+    }
+  }
+  SBase = Addr;
+  Offset = CurDAG->getTargetConstant(0, SL, MVT::i32);
+  Imm = true;
+  return true;
+}
+
+bool AMDGPUDAGToDAGISel::SelectSMRDImm(SDValue Addr, SDValue &SBase,
+                                       SDValue &Offset) const {
+  bool Imm;
+  return SelectSMRD(Addr, SBase, Offset, Imm) && Imm;
+}
+
+bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase,
+                                        SDValue &Offset) const {
+  bool Imm;
+  return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm;
+}
+
+bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr,
+                                             SDValue &Offset) const {
+  bool Imm;
+  return SelectSMRDOffset(Addr, Offset, Imm) && Imm;
+}
+
+bool AMDGPUDAGToDAGISel::SelectSMRDBufferSgpr(SDValue Addr,
+                                              SDValue &Offset) const {
+  bool Imm;
+  return SelectSMRDOffset(Addr, Offset, Imm) && !Imm;
+}
+
 // 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() &&
          "addrspacecast only supported with flat address space!");
 

lib/Target/AMDGPU/SIInstrFormats.td

class SMRDe <bits<5> op, bits<1> imm> : Enc32 {
   let Inst{7-0} = offset;
   let Inst{8} = imm;
   let Inst{14-9} = sbase{6-1};
   let Inst{21-15} = sdst;
   let Inst{26-22} = op;
   let Inst{31-27} = 0x18; //encoding
 }
 
+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{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;
   let hasSideEffects = 0;
   let isCodeGenOnly = 0;
   let SALU = 1;

lib/Target/AMDGPU/SIInstrInfo.td

 
 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 SMRDSgpr  : ComplexPattern<i64, 2, "SelectSMRDSgpr">;
+def SMRDBufferImm   : ComplexPattern<i32, 1, "SelectSMRDBufferImm">;
+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">;
 
 //===----------------------------------------------------------------------===//

 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),
     opName#" $dst, $sbase, $offset", []
   >;
 
+  def _IMM_ci : SMRD <
+    (outs dstClass:$dst), (ins baseClass:$sbase, u32imm:$offset),
+    opName#" $dst, $sbase, $offset", []>, SMRD_IMMe_ci <op> {
+    let AssemblerPredicates = [isCI];
+  }
+
   defm _SGPR : SMRD_m <
     op, opName#"_SGPR", 0, (outs dstClass:$dst),
     (ins baseClass:$sbase, SReg_32:$soff),
     opName#" $dst, $sbase, $soff", []
   >;
 }
 
 //===----------------------------------------------------------------------===//

lib/Target/AMDGPU/SIInstructions.td

 >;
 
 //===----------------------------------------------------------------------===//
 // SMRD Patterns
 //===----------------------------------------------------------------------===//
 
 multiclass SMRD_Pattern <SMRD Instr_IMM, SMRD Instr_SGPR, ValueType vt> {
 
-  // 1. SI-CI: Offset as 8bit DWORD immediate
+  // 1. IMM offset
   def : Pat <
-    (constant_load (add i64:$sbase, (i64 IMM8bitDWORD:$offset))),
-    (vt (Instr_IMM $sbase, (as_dword_i32imm $offset)))
+    (constant_load (SMRDImm i64:$sbase, i32:$offset)),
+    (vt (Instr_IMM $sbase, $offset))
   >;
 
-  // 2. Offset loaded in an 32bit SGPR
-  def : Pat <
-    (constant_load (add i64:$sbase, (i64 IMM32bit:$offset))),
-    (vt (Instr_SGPR $sbase, (S_MOV_B32 (i32 (as_i32imm $offset)))))
-  >;
-
-  // 3. No offset at all
+  // 2. SGPR offset
   def : Pat <
-    (constant_load i64:$sbase),
-    (vt (Instr_IMM $sbase, 0))
+    (constant_load (SMRDSgpr i64:$sbase, i32:$offset)),
+    (vt (Instr_SGPR $sbase, $offset))
   >;
 }
 
-multiclass SMRD_Pattern_vi <SMRD Instr_IMM, SMRD Instr_SGPR, ValueType vt> {
-
-  // 1. VI: Offset as 20bit immediate in bytes
-  def : Pat <
-    (constant_load (add i64:$sbase, (i64 IMM20bit:$offset))),
-    (vt (Instr_IMM $sbase, (as_i32imm $offset)))
-  >;
-
-  // 2. Offset loaded in an 32bit SGPR
-  def : Pat <
-    (constant_load (add i64:$sbase, (i64 IMM32bit:$offset))),
-    (vt (Instr_SGPR $sbase, (S_MOV_B32 (i32 (as_i32imm $offset)))))
-  >;
-
-  // 3. No offset at all
-  def : Pat <
-    (constant_load i64:$sbase),
-    (vt (Instr_IMM $sbase, 0))
-  >;
-}
-
-let Predicates = [isSICI] in {
 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>;
-} // End Predicates = [isSICI]
 
-let Predicates = [isVI] in {
-defm : SMRD_Pattern_vi <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, f32>;
-defm : SMRD_Pattern_vi <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, i32>;
-defm : SMRD_Pattern_vi <S_LOAD_DWORDX2_IMM, S_LOAD_DWORDX2_SGPR, v2i32>;
-defm : SMRD_Pattern_vi <S_LOAD_DWORDX4_IMM, S_LOAD_DWORDX4_SGPR, v4i32>;
-defm : SMRD_Pattern_vi <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v32i8>;
-defm : SMRD_Pattern_vi <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v8i32>;
-defm : SMRD_Pattern_vi <S_LOAD_DWORDX16_IMM, S_LOAD_DWORDX16_SGPR, v16i32>;
-} // End Predicates = [isVI]
-
-let Predicates = [isSICI] in {
-
-// 1. Offset as 8bit DWORD immediate
+// 1. Offset as an immediate
 def : Pat <
-  (SIload_constant v4i32:$sbase, IMM8bitDWORD:$offset),
-  (S_BUFFER_LOAD_DWORD_IMM $sbase, (as_dword_i32imm $offset))
+  (SIload_constant v4i32:$sbase, (SMRDBufferImm i32:$offset)),
+  (S_BUFFER_LOAD_DWORD_IMM $sbase, $offset)
 >;
 
-} // End Predicates = [isSICI]
-
 // 2. Offset loaded in an 32bit SGPR
 def : Pat <
-  (SIload_constant v4i32:$sbase, imm:$offset),
-  (S_BUFFER_LOAD_DWORD_SGPR $sbase, (S_MOV_B32 imm:$offset))
+  (SIload_constant v4i32:$sbase, (SMRDBufferSgpr i32:$offset)),
+  (S_BUFFER_LOAD_DWORD_SGPR $sbase, $offset)
 >;
 
 //===----------------------------------------------------------------------===//
 // SOP1 Patterns
 //===----------------------------------------------------------------------===//
 
 def : Pat <
   (i64 (ctpop i64:$src)),

test/CodeGen/AMDGPU/smrd.ll

 define void @smrd3(i32 addrspace(1)* %out, i32 addrspace(2)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(2)* %ptr, i64 4294967296 ; 2 ^ 32
   %1 = load i32, i32 addrspace(2)* %0
   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]]
+; 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]]
+; 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
 ; 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

 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]]
+; 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]]
+; 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
+}
+
 ; Function Attrs: nounwind readnone
 declare float @llvm.SI.load.const(<16 x i8>, i32) #1
 
 declare void @llvm.SI.export(i32, i32, i32, i32, i32, float, float, float, float)
 
 attributes #0 = { "ShaderType"="0" }
 attributes #1 = { nounwind readnone }