[AMDGPU] SDWA: add support for GFX9 in peephole pass

Summary:
Added support based on merged SDWA pseudo instructions. Now peephole allow one scalar operand, omod and clamp modifiers.
Added several subtarget features for GFX9 SDWA.
This diff also contains changes from D34026.
Depends D34026

Reviewers: vpykhtin, rampitec, arsenm

Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye

Differential Revision: https://reviews.llvm.org/D34241

llvm-svn: 305986

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -238,6 +238,36 @@ def FeatureSDWA : SubtargetFeature<"sdwa",
   "Support SDWA (Sub-DWORD Addressing) extension"
 >;
 
+def FeatureSDWAOmod : SubtargetFeature<"sdwa-omod",
+  "HasSDWAOmod",
+  "true",
+  "Support OMod with SDWA (Sub-DWORD Addressing) extension"
+>;
+
+def FeatureSDWAScalar : SubtargetFeature<"sdwa-scalar",
+  "HasSDWAScalar",
+  "true",
+  "Support scalar register with SDWA (Sub-DWORD Addressing) extension"
+>;
+
+def FeatureSDWASdst : SubtargetFeature<"sdwa-sdst",
+  "HasSDWASdst",
+  "true",
+  "Support scalar dst for VOPC with SDWA (Sub-DWORD Addressing) extension"
+>;
+
+def FeatureSDWAMac : SubtargetFeature<"sdwa-mav",
+  "HasSDWAMac",
+  "true",
+  "Support v_mac_f32/f16 with SDWA (Sub-DWORD Addressing) extension"
+>;
+
+def FeatureSDWAClampVOPC : SubtargetFeature<"sdwa-clamp-vopc",
+  "HasSDWAClampVOPC",
+  "true",
+  "Support clamp for VOPC with SDWA (Sub-DWORD Addressing) extension"
+>;
+
 def FeatureDPP : SubtargetFeature<"dpp",
   "HasDPP",
   "true",
@@ -421,8 +451,8 @@ def FeatureVolcanicIslands : SubtargetFeatureGeneration<"VOLCANIC_ISLANDS",
    FeatureWavefrontSize64, FeatureFlatAddressSpace, FeatureGCN,
    FeatureGCN3Encoding, FeatureCIInsts, Feature16BitInsts,
    FeatureSMemRealTime, FeatureVGPRIndexMode, FeatureMovrel,
-   FeatureScalarStores, FeatureInv2PiInlineImm, FeatureSDWA,
-   FeatureDPP
+   FeatureScalarStores, FeatureInv2PiInlineImm,
+   FeatureSDWA, FeatureSDWAClampVOPC, FeatureSDWAMac, FeatureDPP
   ]
 >;
 
@@ -432,7 +462,8 @@ def FeatureGFX9 : SubtargetFeatureGeneration<"GFX9",
    FeatureGCN3Encoding, FeatureCIInsts, Feature16BitInsts,
    FeatureSMemRealTime, FeatureScalarStores, FeatureInv2PiInlineImm,
    FeatureApertureRegs, FeatureGFX9Insts, FeatureVOP3P, FeatureVGPRIndexMode,
-   FeatureFastFMAF32, FeatureSDWA, FeatureDPP,
+   FeatureFastFMAF32, FeatureDPP,
+   FeatureSDWA, FeatureSDWAOmod, FeatureSDWAScalar, FeatureSDWASdst,
    FeatureFlatInstOffsets, FeatureFlatGlobalInsts, FeatureFlatScratchInsts
   ]
 >;

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp

@@ -124,6 +124,11 @@ AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     HasScalarStores(false),
     HasInv2PiInlineImm(false),
     HasSDWA(false),
+    HasSDWAOmod(false),
+    HasSDWAScalar(false),
+    HasSDWASdst(false),
+    HasSDWAMac(false),
+    HasSDWAClampVOPC(false),
     HasDPP(false),
     FlatAddressSpace(false),
     FlatInstOffsets(false),

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

@@ -149,6 +149,11 @@ class AMDGPUSubtarget : public AMDGPUGenSubtargetInfo {
   bool HasScalarStores;
   bool HasInv2PiInlineImm;
   bool HasSDWA;
+  bool HasSDWAOmod;
+  bool HasSDWAScalar;
+  bool HasSDWASdst;
+  bool HasSDWAMac;
+  bool HasSDWAClampVOPC;
   bool HasDPP;
   bool FlatAddressSpace;
   bool FlatInstOffsets;
@@ -431,6 +436,26 @@ class AMDGPUSubtarget : public AMDGPUGenSubtargetInfo {
     return HasSDWA;
   }
 
+  bool hasSDWAOmod() const {
+    return HasSDWAOmod;
+  }
+
+  bool hasSDWAScalar() const {
+    return HasSDWAScalar;
+  }
+
+  bool hasSDWASdst() const {
+    return HasSDWASdst;
+  }
+
+  bool hasSDWAMac() const {
+    return HasSDWAMac;
+  }
+
+  bool hasSDWAClampVOPC() const {
+    return HasSDWAClampVOPC;
+  }
+
   /// \brief Returns the offset in bytes from the start of the input buffer
   ///        of the first explicit kernel argument.
   unsigned getExplicitKernelArgOffset(const MachineFunction &MF) const {

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

@@ -2454,7 +2454,7 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
         continue;
       const MachineOperand &MO = MI.getOperand(OpIdx);
 
-      if (AMDGPU::isVI(ST)) {
+      if (!ST.hasSDWAScalar()) {
         // Only VGPRS on VI
         if (!MO.isReg() || !RI.hasVGPRs(RI.getRegClassForReg(MRI, MO.getReg()))) {
           ErrInfo = "Only VGPRs allowed as operands in SDWA instructions on VI";
@@ -2469,7 +2469,7 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
       }
     }
 
-    if (AMDGPU::isVI(ST)) {
+    if (!ST.hasSDWAOmod()) {
       // No omod allowed on VI
       const MachineOperand *OMod = getNamedOperand(MI, AMDGPU::OpName::omod);
       if (OMod != nullptr &&
@@ -2481,14 +2481,14 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
 
     uint16_t BasicOpcode = AMDGPU::getBasicFromSDWAOp(Opcode);
     if (isVOPC(BasicOpcode)) {
-      if (AMDGPU::isVI(ST) && DstIdx != -1) {
+      if (!ST.hasSDWASdst() && DstIdx != -1) {
         // Only vcc allowed as dst on VI for VOPC
         const MachineOperand &Dst = MI.getOperand(DstIdx);
         if (!Dst.isReg() || Dst.getReg() != AMDGPU::VCC) {
           ErrInfo = "Only VCC allowed as dst in SDWA instructions on VI";
           return false;
         }
-      } else if (AMDGPU::isGFX9(ST)) {
+      } else if (!ST.hasSDWAClampVOPC()) {
         // No clamp allowed on GFX9 for VOPC
         const MachineOperand *Clamp = getNamedOperand(MI, AMDGPU::OpName::clamp);
         if (Clamp != nullptr &&

llvm/lib/Target/AMDGPU/SIPeepholeSDWA.cpp

@@ -67,9 +67,9 @@ class SIPeepholeSDWA : public MachineFunctionPass {
 
   bool runOnMachineFunction(MachineFunction &MF) override;
   void matchSDWAOperands(MachineFunction &MF);
-  bool isConvertibleToSDWA(const MachineInstr &MI) const;
+  bool isConvertibleToSDWA(const MachineInstr &MI, const SISubtarget &ST) const;
   bool convertToSDWA(MachineInstr &MI, const SDWAOperandsVector &SDWAOperands);
-  void legalizeScalarOperands(MachineInstr &MI) const;
+  void legalizeScalarOperands(MachineInstr &MI, const SISubtarget &ST) const;
 
   StringRef getPassName() const override { return "SI Peephole SDWA"; }
 
@@ -607,24 +607,38 @@ void SIPeepholeSDWA::matchSDWAOperands(MachineFunction &MF) {
   }
 }
 
-bool SIPeepholeSDWA::isConvertibleToSDWA(const MachineInstr &MI) const {
+bool SIPeepholeSDWA::isConvertibleToSDWA(const MachineInstr &MI,
+                                         const SISubtarget &ST) const {
   // Check if this instruction has opcode that supports SDWA
-  unsigned Opc = MI.getOpcode();
-  if (AMDGPU::getSDWAOp(Opc) != -1)
-    return true;
-  int Opc32 = AMDGPU::getVOPe32(Opc);
-  if (Opc32 != -1 && AMDGPU::getSDWAOp(Opc32) != -1) {
-    if (TII->hasModifiersSet(MI, AMDGPU::OpName::omod))
-      return false;
+  int Opc = MI.getOpcode();
+  if (AMDGPU::getSDWAOp(Opc) == -1)
+    Opc = AMDGPU::getVOPe32(Opc);
+
+  if (Opc == -1 || AMDGPU::getSDWAOp(Opc) == -1)
+    return false;
+
+  if (!ST.hasSDWAOmod() && TII->hasModifiersSet(MI, AMDGPU::OpName::omod))
+    return false;
 
-    if (TII->isVOPC(Opc)) {
+  if (TII->isVOPC(Opc)) {
+    if (!ST.hasSDWASdst()) {
       const MachineOperand *SDst = TII->getNamedOperand(MI, AMDGPU::OpName::sdst);
-      return SDst && SDst->getReg() == AMDGPU::VCC;
-    } else {
-      return !TII->getNamedOperand(MI, AMDGPU::OpName::sdst);
+      if (SDst && SDst->getReg() != AMDGPU::VCC)
+        return false;
     }
+
+    if (!ST.hasSDWAClampVOPC() && TII->hasModifiersSet(MI, AMDGPU::OpName::clamp))
+      return false;
+
+  } else if (TII->getNamedOperand(MI, AMDGPU::OpName::sdst)) {
+    return false;
   }
-  return false;
+
+  if (!ST.hasSDWAMac() && (Opc == AMDGPU::V_MAC_F16_e32 ||
+                           Opc == AMDGPU::V_MAC_F32_e32))
+    return false;
+
+  return true;
 }
 
 bool SIPeepholeSDWA::convertToSDWA(MachineInstr &MI,
@@ -690,13 +704,23 @@ bool SIPeepholeSDWA::convertToSDWA(MachineInstr &MI,
     SDWAInst.add(*Src2);
   }
 
-  // Initialize clamp.
-  if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::clamp) != -1)
+  // Copy clamp if present, initialize otherwise
+  assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::clamp) != -1);
+  MachineOperand *Clamp = TII->getNamedOperand(MI, AMDGPU::OpName::clamp);
+  if (Clamp) {
+    SDWAInst.add(*Clamp);
+  } else {
     SDWAInst.addImm(0);
+  }
 
-  // Initialize omod.
-  if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::omod) != -1)
+  // Copy omod if present, initialize otherwise if needed
+  MachineOperand *OMod = TII->getNamedOperand(MI, AMDGPU::OpName::omod);
+  if (OMod) {
+    assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::omod) != -1);
+    SDWAInst.add(*OMod);
+  } else if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::omod) != -1) {
     SDWAInst.addImm(0);
+  }
 
   // Initialize dst_sel and dst_unused if present
   if (Dst) {
@@ -750,16 +774,25 @@ bool SIPeepholeSDWA::convertToSDWA(MachineInstr &MI,
 }
 
 // If an instruction was converted to SDWA it should not have immediates or SGPR
-// operands. Copy its scalar operands into VGPRs.
-void SIPeepholeSDWA::legalizeScalarOperands(MachineInstr &MI) const {
+// operands (allowed one SGPR on GFX9). Copy its scalar operands into VGPRs.
+void SIPeepholeSDWA::legalizeScalarOperands(MachineInstr &MI, const SISubtarget &ST) const {
   const MCInstrDesc &Desc = TII->get(MI.getOpcode());
-  for (unsigned I = 0, E = MI.getNumExplicitOperands(); I != E; ++I) {
-    MachineOperand &Op = MI.getOperand(I);
+  unsigned ConstantBusCount = 0;
+  for (MachineOperand &Op: MI.explicit_uses()) {
     if (!Op.isImm() && !(Op.isReg() && !TRI->isVGPR(*MRI, Op.getReg())))
       continue;
+
+    unsigned I = MI.getOperandNo(&Op);
     if (Desc.OpInfo[I].RegClass == -1 ||
        !TRI->hasVGPRs(TRI->getRegClass(Desc.OpInfo[I].RegClass)))
       continue;
+
+    if (ST.hasSDWAScalar() && ConstantBusCount == 0 && Op.isReg() &&
+        TRI->isSGPRReg(*MRI, Op.getReg())) {
+      ++ConstantBusCount;
+      continue;
+    }
+
     unsigned VGPR = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
     auto Copy = BuildMI(*MI.getParent(), MI.getIterator(), MI.getDebugLoc(),
                         TII->get(AMDGPU::V_MOV_B32_e32), VGPR);
@@ -775,10 +808,8 @@ void SIPeepholeSDWA::legalizeScalarOperands(MachineInstr &MI) const {
 bool SIPeepholeSDWA::runOnMachineFunction(MachineFunction &MF) {
   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
 
-  if (!ST.hasSDWA() ||
-      !AMDGPU::isVI(ST)) { // TODO: Add support for SDWA on gfx9
+  if (!ST.hasSDWA())
     return false;
-  }
 
   MRI = &MF.getRegInfo();
   TRI = ST.getRegisterInfo();
@@ -790,7 +821,7 @@ bool SIPeepholeSDWA::runOnMachineFunction(MachineFunction &MF) {
   for (const auto &OperandPair : SDWAOperands) {
     const auto &Operand = OperandPair.second;
     MachineInstr *PotentialMI = Operand->potentialToConvert(TII);
-    if (PotentialMI && isConvertibleToSDWA(*PotentialMI)) {
+    if (PotentialMI && isConvertibleToSDWA(*PotentialMI, ST)) {
       PotentialMatches[PotentialMI].push_back(Operand.get());
     }
   }
@@ -805,7 +836,7 @@ bool SIPeepholeSDWA::runOnMachineFunction(MachineFunction &MF) {
 
   bool Ret = !ConvertedInstructions.empty();
   while (!ConvertedInstructions.empty())
-    legalizeScalarOperands(*ConvertedInstructions.pop_back_val());
+    legalizeScalarOperands(*ConvertedInstructions.pop_back_val(), ST);
 
   return Ret;
 }

llvm/lib/Target/AMDGPU/VOPInstructions.td

@@ -401,10 +401,6 @@ class VOP_SDWA_Real <VOP_SDWA_Pseudo ps> :
   let Constraints     = ps.Constraints;
   let DisableEncoding = ps.DisableEncoding;
 
-  // string Mnemonic     = ps.Mnemonic;
-  // string AsmOperands  = ps.AsmOperands;
-  // string AsmOperands9 = ps.AsmOperands9;
-
   // Copy relevant pseudo op flags
   let SubtargetPredicate   = ps.SubtargetPredicate;
   let AssemblerPredicate   = ps.AssemblerPredicate;

llvm/test/CodeGen/AMDGPU/fneg.f16.ll

@@ -134,11 +134,10 @@ define amdgpu_kernel void @v_fneg_fold_v2f16(<2 x half> addrspace(1)* %out, <2 x
 ; CI-DAG: v_mul_f32_e32 v{{[0-9]+}}, -4.0, v{{[0-9]+}}
 ; CI-DAG: v_sub_f32_e32 v{{[0-9]+}}, 2.0, v{{[0-9]+}}
 
-; GFX9: v_lshrrev_b32_e32 [[ELT1:v[0-9]+]], 16, [[VAL]]
 ; GFX89-DAG: v_mul_f16_e32 v{{[0-9]+}}, -4.0, [[VAL]]
-; GFX9-DAG: v_sub_f16_e32 v{{[0-9]+}}, 2.0, [[ELT1]]
-; VI-DAG: v_mov_b32_e32 [[CONST2:v[0-9]+]], 0x4000
-; VI-DAG: v_sub_f16_sdwa v{{[0-9]+}}, [[CONST2]], [[VAL]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX89-DAG: v_mov_b32_e32 [[CONST2:v[0-9]+]], 0x4000
+; GFX89-DAG: v_sub_f16_sdwa v{{[0-9]+}}, [[CONST2]], [[VAL]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
 define amdgpu_kernel void @v_extract_fneg_fold_v2f16(<2 x half> addrspace(1)* %in) #0 {
   %val = load <2 x half>, <2 x half> addrspace(1)* %in
   %fneg = fsub <2 x half> <half -0.0, half -0.0>, %val

llvm/test/CodeGen/AMDGPU/fpext.f16.ll

@@ -1,6 +1,6 @@
-; RUN: llc -march=amdgcn -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=GCN -check-prefix=SI -check-prefix=SIVI -check-prefix=SIGFX9 %s
+; RUN: llc -march=amdgcn -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=GCN -check-prefix=SI -check-prefix=SIVI %s
 ; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=GFX89 %s
-; RUN: llc -march=amdgcn -mcpu=gfx901 -mattr=-flat-for-global -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -check-prefix=GFX89 -check-prefix=SIGFX9 %s
+; RUN: llc -march=amdgcn -mcpu=gfx901 -mattr=-flat-for-global -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -check-prefix=GFX89 %s
 
 ; GCN-LABEL: {{^}}fpext_f16_to_f32
 ; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
@@ -35,11 +35,10 @@ entry:
 
 ; GCN-LABEL: {{^}}fpext_v2f16_to_v2f32
 ; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GFX9-DAG:  v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
 ; GCN-DAG: v_cvt_f32_f16_e32 v[[R_F32_0:[0-9]+]], v[[A_V2_F16]]
 ; SI:  v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SIGFX9: v_cvt_f32_f16_e32 v[[R_F32_1:[0-9]+]], v[[A_F16_1]]
-; VI: v_cvt_f32_f16_sdwa v[[R_F32_1:[0-9]+]], v[[A_V2_F16]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
+; SI: v_cvt_f32_f16_e32 v[[R_F32_1:[0-9]+]], v[[A_F16_1]]
+; GFX89: v_cvt_f32_f16_sdwa v[[R_F32_1:[0-9]+]], v[[A_V2_F16]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
 ; GCN: buffer_store_dwordx2 v{{\[}}[[R_F32_0]]:[[R_F32_1]]{{\]}}
 ; GCN: s_endpgm
 
@@ -55,9 +54,9 @@ entry:
 
 ; GCN-LABEL: {{^}}fpext_v2f16_to_v2f64
 ; GCN: buffer_load_dword
-; SIGFX9-DAG: v_lshrrev_b32_e32
-; SIGFX9-DAG: v_cvt_f32_f16_e32
-; VI: v_cvt_f32_f16_sdwa
+; SI-DAG: v_lshrrev_b32_e32
+; SI-DAG: v_cvt_f32_f16_e32
+; GFX89: v_cvt_f32_f16_sdwa
 ; GCN: v_cvt_f32_f16_e32
 
 ; GCN: v_cvt_f64_f32_e32

llvm/test/CodeGen/AMDGPU/llvm.rint.f16.ll

@@ -42,13 +42,13 @@ entry:
 ; VI:     v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_1]], v[[R_F16_0]]
 
 ; GFX9: v_rndne_f16_e32 v[[R_F16_0:[0-9]+]], v[[A_V2_F16]]
-; GFX9: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; GFX9: v_rndne_f16_e32 v[[R_F16_1:[0-9]+]], v[[A_F16_1]]
+; GFX9: v_rndne_f16_sdwa v[[R_F16_1:[0-9]+]], v[[A_V2_F16]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
 ; GFX9: v_and_b32_e32 v[[R_F16_LO:[0-9]+]], 0xffff, v[[R_F16_0]]
 ; GFX9: v_lshl_or_b32 v[[R_V2_F16:[0-9]+]], v[[R_F16_1]], 16, v[[R_F16_LO]]
 
 ; GCN: buffer_store_dword v[[R_V2_F16]]
 ; GCN: s_endpgm
+
 define amdgpu_kernel void @rint_v2f16(
     <2 x half> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a) {