1 files changed, 136 insertions, 10 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index bc0fd8d..9017f4f 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -874,13 +874,15 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
 
   setOperationAction({ISD::SMULO, ISD::UMULO}, MVT::i64, Custom);
 
-  if (Subtarget->hasScalarSMulU64())
+  if (Subtarget->hasVectorMulU64())
+    setOperationAction(ISD::MUL, MVT::i64, Legal);
+  else if (Subtarget->hasScalarSMulU64())
     setOperationAction(ISD::MUL, MVT::i64, Custom);
 
   if (Subtarget->hasMad64_32())
     setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, MVT::i32, Custom);
 
-  if (Subtarget->hasPrefetch() && Subtarget->hasSafeSmemPrefetch())
+  if (Subtarget->hasSafeSmemPrefetch() || Subtarget->hasVmemPrefInsts())
     setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
 
   if (Subtarget->hasIEEEMinimumMaximumInsts()) {
@@ -944,6 +946,12 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::BUILD_VECTOR, MVT::v2bf16, Legal);
   }
 
+  if (Subtarget->hasBF16PackedInsts()) {
+    setOperationAction(
+        {ISD::FADD, ISD::FMUL, ISD::FMINNUM, ISD::FMAXNUM, ISD::FMA},
+        MVT::v2bf16, Legal);
+  }
+
   if (Subtarget->hasBF16TransInsts()) {
     setOperationAction({ISD::FEXP2, ISD::FLOG2, ISD::FSQRT}, MVT::bf16, Legal);
   }
@@ -1053,10 +1061,12 @@ ArrayRef<MCPhysReg> SITargetLowering::getRoundingControlRegisters() const {
 // where this is OK to use.
 bool SITargetLowering::isFPExtFoldable(const SelectionDAG &DAG, unsigned Opcode,
                                        EVT DestVT, EVT SrcVT) const {
-  return ((Opcode == ISD::FMAD && Subtarget->hasMadMixInsts()) ||
-          (Opcode == ISD::FMA && Subtarget->hasFmaMixInsts())) &&
-         DestVT.getScalarType() == MVT::f32 &&
-         SrcVT.getScalarType() == MVT::f16 &&
+  return DestVT.getScalarType() == MVT::f32 &&
+         ((((Opcode == ISD::FMAD && Subtarget->hasMadMixInsts()) ||
+            (Opcode == ISD::FMA && Subtarget->hasFmaMixInsts())) &&
+           SrcVT.getScalarType() == MVT::f16) ||
+          (Opcode == ISD::FMA && Subtarget->hasFmaMixBF16Insts() &&
+           SrcVT.getScalarType() == MVT::bf16)) &&
          // TODO: This probably only requires no input flushing?
          denormalModeIsFlushAllF32(DAG.getMachineFunction());
 }
@@ -1467,6 +1477,12 @@ bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
                   MachineMemOperand::MOVolatile;
     return true;
   }
+  case Intrinsic::amdgcn_flat_load_monitor_b32:
+  case Intrinsic::amdgcn_flat_load_monitor_b64:
+  case Intrinsic::amdgcn_flat_load_monitor_b128:
+  case Intrinsic::amdgcn_global_load_monitor_b32:
+  case Intrinsic::amdgcn_global_load_monitor_b64:
+  case Intrinsic::amdgcn_global_load_monitor_b128:
   case Intrinsic::amdgcn_ds_load_tr6_b96:
   case Intrinsic::amdgcn_ds_load_tr4_b64:
   case Intrinsic::amdgcn_ds_load_tr8_b64:
@@ -1540,7 +1556,9 @@ bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
     Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
     return true;
   }
-  case Intrinsic::amdgcn_s_prefetch_data: {
+  case Intrinsic::amdgcn_s_prefetch_data:
+  case Intrinsic::amdgcn_flat_prefetch:
+  case Intrinsic::amdgcn_global_prefetch: {
     Info.opc = ISD::INTRINSIC_VOID;
     Info.memVT = EVT::getIntegerVT(CI.getContext(), 8);
     Info.ptrVal = CI.getArgOperand(0);
@@ -1591,10 +1609,16 @@ bool SITargetLowering::getAddrModeArguments(const IntrinsicInst *II,
   case Intrinsic::amdgcn_ds_atomic_barrier_arrive_rtn_b64:
   case Intrinsic::amdgcn_flat_atomic_fmax_num:
   case Intrinsic::amdgcn_flat_atomic_fmin_num:
+  case Intrinsic::amdgcn_flat_load_monitor_b128:
+  case Intrinsic::amdgcn_flat_load_monitor_b32:
+  case Intrinsic::amdgcn_flat_load_monitor_b64:
   case Intrinsic::amdgcn_global_atomic_csub:
   case Intrinsic::amdgcn_global_atomic_fmax_num:
   case Intrinsic::amdgcn_global_atomic_fmin_num:
   case Intrinsic::amdgcn_global_atomic_ordered_add_b64:
+  case Intrinsic::amdgcn_global_load_monitor_b128:
+  case Intrinsic::amdgcn_global_load_monitor_b32:
+  case Intrinsic::amdgcn_global_load_monitor_b64:
   case Intrinsic::amdgcn_global_load_tr_b64:
   case Intrinsic::amdgcn_global_load_tr_b128:
   case Intrinsic::amdgcn_global_load_tr4_b64:
@@ -4432,19 +4456,28 @@ SDValue SITargetLowering::lowerSET_ROUNDING(SDValue Op,
 }
 
 SDValue SITargetLowering::lowerPREFETCH(SDValue Op, SelectionDAG &DAG) const {
-  if (Op->isDivergent())
+  if (Op->isDivergent() &&
+      (!Subtarget->hasVmemPrefInsts() || !Op.getConstantOperandVal(4)))
+    // Cannot do I$ prefetch with divergent pointer.
     return SDValue();
 
   switch (cast<MemSDNode>(Op)->getAddressSpace()) {
   case AMDGPUAS::FLAT_ADDRESS:
   case AMDGPUAS::GLOBAL_ADDRESS:
   case AMDGPUAS::CONSTANT_ADDRESS:
-  case AMDGPUAS::CONSTANT_ADDRESS_32BIT:
     break;
+  case AMDGPUAS::CONSTANT_ADDRESS_32BIT:
+    if (Subtarget->hasSafeSmemPrefetch())
+      break;
+    [[fallthrough]];
   default:
     return SDValue();
   }
 
+  // I$ prefetch
+  if (!Subtarget->hasSafeSmemPrefetch() && !Op.getConstantOperandVal(4))
+    return SDValue();
+
   return Op;
 }
 
@@ -5415,6 +5448,19 @@ SITargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
     MachineOperand &Src0 = MI.getOperand(1);
     MachineOperand &Src1 = MI.getOperand(2);
 
+    if (ST.hasAddSubU64Insts()) {
+      auto I = BuildMI(*BB, MI, DL,
+                       TII->get(IsAdd ? AMDGPU::V_ADD_U64_e64
+                                      : AMDGPU::V_SUB_U64_e64),
+                       Dest.getReg())
+                   .add(Src0)
+                   .add(Src1)
+                   .addImm(0); // clamp
+      TII->legalizeOperands(*I);
+      MI.eraseFromParent();
+      return BB;
+    }
+
     if (IsAdd && ST.hasLshlAddU64Inst()) {
       auto Add = BuildMI(*BB, MI, DL, TII->get(AMDGPU::V_LSHL_ADD_U64_e64),
                          Dest.getReg())
@@ -13633,6 +13679,7 @@ bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
     case Intrinsic::amdgcn_rcp_legacy:
     case Intrinsic::amdgcn_rsq_legacy:
     case Intrinsic::amdgcn_trig_preop:
+    case Intrinsic::amdgcn_tanh:
     case Intrinsic::amdgcn_log:
     case Intrinsic::amdgcn_exp2:
     case Intrinsic::amdgcn_sqrt:
@@ -14046,7 +14093,8 @@ static bool supportsMin3Max3(const GCNSubtarget &Subtarget, unsigned Opc,
   case ISD::FMAXIMUMNUM:
   case AMDGPUISD::FMIN_LEGACY:
   case AMDGPUISD::FMAX_LEGACY:
-    return (VT == MVT::f32) || (VT == MVT::f16 && Subtarget.hasMin3Max3_16());
+    return (VT == MVT::f32) || (VT == MVT::f16 && Subtarget.hasMin3Max3_16()) ||
+           (VT == MVT::v2f16 && Subtarget.hasMin3Max3PKF16());
   case ISD::FMINIMUM:
   case ISD::FMAXIMUM:
     return (VT == MVT::f32 && Subtarget.hasMinimum3Maximum3F32()) ||
@@ -14131,6 +14179,8 @@ SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
         Op0.getOpcode() == AMDGPUISD::FMAX_LEGACY)) &&
       (VT == MVT::f32 || VT == MVT::f64 ||
        (VT == MVT::f16 && Subtarget->has16BitInsts()) ||
+       (VT == MVT::bf16 && Subtarget->hasBF16PackedInsts()) ||
+       (VT == MVT::v2bf16 && Subtarget->hasBF16PackedInsts()) ||
        (VT == MVT::v2f16 && Subtarget->hasVOP3PInsts())) &&
       Op0.hasOneUse()) {
     if (SDValue Res = performFPMed3ImmCombine(DAG, SDLoc(N), Op0, Op1))
@@ -15846,6 +15896,78 @@ SDValue SITargetLowering::performClampCombine(SDNode *N,
   return SDValue(CSrc, 0);
 }
 
+SDValue SITargetLowering::performSelectCombine(SDNode *N,
+                                               DAGCombinerInfo &DCI) const {
+
+  // Try to fold CMP + SELECT patterns with shared constants (both FP and
+  // integer).
+  // Detect when CMP and SELECT use the same constant and fold them to avoid
+  // loading the constant twice. Specifically handles patterns like:
+  // %cmp = icmp eq i32 %val, 4242
+  // %sel = select i1 %cmp, i32 4242, i32 %other
+  // It can be optimized to reuse %val instead of 4242 in select.
+  SDValue Cond = N->getOperand(0);
+  SDValue TrueVal = N->getOperand(1);
+  SDValue FalseVal = N->getOperand(2);
+
+  // Check if condition is a comparison.
+  if (Cond.getOpcode() != ISD::SETCC)
+    return SDValue();
+
+  SDValue LHS = Cond.getOperand(0);
+  SDValue RHS = Cond.getOperand(1);
+  ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+
+  bool isFloatingPoint = LHS.getValueType().isFloatingPoint();
+  bool isInteger = LHS.getValueType().isInteger();
+
+  // Handle simple floating-point and integer types only.
+  if (!isFloatingPoint && !isInteger)
+    return SDValue();
+
+  bool isEquality = CC == (isFloatingPoint ? ISD::SETOEQ : ISD::SETEQ);
+  bool isNonEquality = CC == (isFloatingPoint ? ISD::SETONE : ISD::SETNE);
+  if (!isEquality && !isNonEquality)
+    return SDValue();
+
+  SDValue ArgVal, ConstVal;
+  if ((isFloatingPoint && isa<ConstantFPSDNode>(RHS)) ||
+      (isInteger && isa<ConstantSDNode>(RHS))) {
+    ConstVal = RHS;
+    ArgVal = LHS;
+  } else if ((isFloatingPoint && isa<ConstantFPSDNode>(LHS)) ||
+             (isInteger && isa<ConstantSDNode>(LHS))) {
+    ConstVal = LHS;
+    ArgVal = RHS;
+  } else {
+    return SDValue();
+  }
+
+  // Skip optimization for inlinable immediates.
+  if (isFloatingPoint) {
+    const APFloat &Val = cast<ConstantFPSDNode>(ConstVal)->getValueAPF();
+    if (!Val.isNormal() || Subtarget->getInstrInfo()->isInlineConstant(Val))
+      return SDValue();
+  } else {
+    if (AMDGPU::isInlinableIntLiteral(
+            cast<ConstantSDNode>(ConstVal)->getSExtValue()))
+      return SDValue();
+  }
+
+  // For equality and non-equality comparisons, patterns:
+  // select (setcc x, const), const, y -> select (setcc x, const), x, y
+  // select (setccinv x, const), y, const -> select (setccinv x, const), y, x
+  if (!(isEquality && TrueVal == ConstVal) &&
+      !(isNonEquality && FalseVal == ConstVal))
+    return SDValue();
+
+  SDValue SelectLHS = (isEquality && TrueVal == ConstVal) ? ArgVal : TrueVal;
+  SDValue SelectRHS =
+      (isNonEquality && FalseVal == ConstVal) ? ArgVal : FalseVal;
+  return DCI.DAG.getNode(ISD::SELECT, SDLoc(N), N->getValueType(0), Cond,
+                         SelectLHS, SelectRHS);
+}
+
 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
                                             DAGCombinerInfo &DCI) const {
   switch (N->getOpcode()) {
@@ -15894,6 +16016,10 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
     return performFMulCombine(N, DCI);
   case ISD::SETCC:
     return performSetCCCombine(N, DCI);
+  case ISD::SELECT:
+    if (auto Res = performSelectCombine(N, DCI))
+      return Res;
+    break;
   case ISD::FMAXNUM:
   case ISD::FMINNUM:
   case ISD::FMAXNUM_IEEE: