AMDGPU: Remove ds atomic fadd intrinsics (#95396)

These have been replaced with atomicrmw fadd

Author: arsenm

clang/lib/CodeGen/CGBuiltin.cpp

@@ -19105,7 +19105,7 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
       ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(2)),
                               EmitScalarExpr(E->getArg(3)), AO, SSID);
     } else {
-      // The ds_fadd_* builtins do not have syncscope/order arguments.
+      // The ds_atomic_fadd_* builtins do not have syncscope/order arguments.
       SSID = llvm::SyncScope::System;
       AO = AtomicOrdering::SequentiallyConsistent;
 

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

@@ -571,7 +571,6 @@ def int_amdgcn_ds_ordered_swap : AMDGPUDSOrderedIntrinsic;
 def int_amdgcn_ds_append : AMDGPUDSAppendConsumedIntrinsic;
 def int_amdgcn_ds_consume : AMDGPUDSAppendConsumedIntrinsic;
 
-def int_amdgcn_ds_fadd : AMDGPULDSIntrin;
 def int_amdgcn_ds_fmin : AMDGPULDSIntrin;
 def int_amdgcn_ds_fmax : AMDGPULDSIntrin;
 
@@ -2930,10 +2929,6 @@ multiclass AMDGPUMFp8SmfmacIntrinsic<LLVMType DestTy> {
 // bf16 atomics use v2i16 argument since there is no bf16 data type in the llvm.
 def int_amdgcn_global_atomic_fadd_v2bf16 : AMDGPUAtomicRtn<llvm_v2i16_ty>;
 def int_amdgcn_flat_atomic_fadd_v2bf16   : AMDGPUAtomicRtn<llvm_v2i16_ty>;
-def int_amdgcn_ds_fadd_v2bf16 : DefaultAttrsIntrinsic<
-    [llvm_v2i16_ty],
-    [LLVMQualPointerType<3>, llvm_v2i16_ty],
-    [IntrArgMemOnly, NoCapture<ArgIndex<0>>]>;
 
 defset list<Intrinsic> AMDGPUMFMAIntrinsics940 = {
 def int_amdgcn_mfma_i32_16x16x32_i8     : AMDGPUMfmaIntrinsic<llvm_v4i32_ty,  llvm_i64_ty>;

llvm/lib/IR/AutoUpgrade.cpp

@@ -1033,6 +1033,12 @@ static bool upgradeIntrinsicFunction1(Function *F, Function *&NewFn,
         break; // No other 'amdgcn.atomic.*'
       }
 
+      if (Name.starts_with("ds.fadd")) {
+        // Replaced with atomicrmw fadd, so there's no new declaration.
+        NewFn = nullptr;
+        return true;
+      }
+
       if (Name.starts_with("ldexp.")) {
         // Target specific intrinsic became redundant
         NewFn = Intrinsic::getDeclaration(
@@ -2331,40 +2337,74 @@ static Value *upgradeARMIntrinsicCall(StringRef Name, CallBase *CI, Function *F,
   llvm_unreachable("Unknown function for ARM CallBase upgrade.");
 }
 
+// These are expected to have the arguments:
+// atomic.intrin (ptr, rmw_value, ordering, scope, isVolatile)
+//
+// Except for int_amdgcn_ds_fadd_v2bf16 which only has (ptr, rmw_value).
+//
 static Value *upgradeAMDGCNIntrinsicCall(StringRef Name, CallBase *CI,
                                          Function *F, IRBuilder<> &Builder) {
-  const bool IsInc = Name.starts_with("atomic.inc.");
-  if (IsInc || Name.starts_with("atomic.dec.")) {
-    if (CI->getNumOperands() != 6) // Malformed bitcode.
-      return nullptr;
+  AtomicRMWInst::BinOp RMWOp =
+      StringSwitch<AtomicRMWInst::BinOp>(Name)
+          .StartsWith("ds.fadd", AtomicRMWInst::FAdd)
+          .StartsWith("atomic.inc.", AtomicRMWInst::UIncWrap)
+          .StartsWith("atomic.dec.", AtomicRMWInst::UDecWrap);
+
+  unsigned NumOperands = CI->getNumOperands();
+  if (NumOperands < 3) // Malformed bitcode.
+    return nullptr;
 
-    AtomicRMWInst::BinOp RMWOp =
-        IsInc ? AtomicRMWInst::UIncWrap : AtomicRMWInst::UDecWrap;
+  Value *Ptr = CI->getArgOperand(0);
+  if (!isa<PointerType>(Ptr->getType())) // Malformed.
+    return nullptr;
 
-    Value *Ptr = CI->getArgOperand(0);
-    Value *Val = CI->getArgOperand(1);
-    ConstantInt *OrderArg = dyn_cast<ConstantInt>(CI->getArgOperand(2));
+  Value *Val = CI->getArgOperand(1);
+  if (Val->getType() != CI->getType()) // Malformed.
+    return nullptr;
+
+  ConstantInt *OrderArg = nullptr;
+  bool IsVolatile = false;
+
+  // These should have 5 arguments (plus the callee). A separate version of the
+  // ds_fadd intrinsic was defined for bf16 which was missing arguments.
+  if (NumOperands > 3)
+    OrderArg = dyn_cast<ConstantInt>(CI->getArgOperand(2));
+
+  // Ignore scope argument at 3
+
+  if (NumOperands > 5) {
     ConstantInt *VolatileArg = dyn_cast<ConstantInt>(CI->getArgOperand(4));
+    IsVolatile = !VolatileArg || !VolatileArg->isZero();
+  }
 
-    AtomicOrdering Order = AtomicOrdering::SequentiallyConsistent;
-    if (OrderArg && isValidAtomicOrdering(OrderArg->getZExtValue()))
-      Order = static_cast<AtomicOrdering>(OrderArg->getZExtValue());
-    if (Order == AtomicOrdering::NotAtomic ||
-        Order == AtomicOrdering::Unordered)
-      Order = AtomicOrdering::SequentiallyConsistent;
-
-    // The scope argument never really worked correctly. Use agent as the most
-    // conservative option which should still always produce the instruction.
-    SyncScope::ID SSID = F->getContext().getOrInsertSyncScopeID("agent");
-    AtomicRMWInst *RMW =
-        Builder.CreateAtomicRMW(RMWOp, Ptr, Val, std::nullopt, Order, SSID);
-
-    if (!VolatileArg || !VolatileArg->isZero())
-      RMW->setVolatile(true);
-    return RMW;
+  AtomicOrdering Order = AtomicOrdering::SequentiallyConsistent;
+  if (OrderArg && isValidAtomicOrdering(OrderArg->getZExtValue()))
+    Order = static_cast<AtomicOrdering>(OrderArg->getZExtValue());
+  if (Order == AtomicOrdering::NotAtomic || Order == AtomicOrdering::Unordered)
+    Order = AtomicOrdering::SequentiallyConsistent;
+
+  LLVMContext &Ctx = F->getContext();
+
+  // Handle the v2bf16 intrinsic which used <2 x i16> instead of <2 x bfloat>
+  Type *RetTy = CI->getType();
+  if (VectorType *VT = dyn_cast<VectorType>(RetTy)) {
+    if (VT->getElementType()->isIntegerTy(16)) {
+      VectorType *AsBF16 =
+          VectorType::get(Type::getBFloatTy(Ctx), VT->getElementCount());
+      Val = Builder.CreateBitCast(Val, AsBF16);
+    }
   }
 
-  llvm_unreachable("Unknown function for AMDGPU intrinsic upgrade.");
+  // The scope argument never really worked correctly. Use agent as the most
+  // conservative option which should still always produce the instruction.
+  SyncScope::ID SSID = Ctx.getOrInsertSyncScopeID("agent");
+  AtomicRMWInst *RMW =
+      Builder.CreateAtomicRMW(RMWOp, Ptr, Val, std::nullopt, Order, SSID);
+
+  if (IsVolatile)
+    RMW->setVolatile(true);
+
+  return Builder.CreateBitCast(RMW, RetTy);
 }
 
 /// Helper to unwrap intrinsic call MetadataAsValue operands.

llvm/lib/Target/AMDGPU/AMDGPUInstructions.td

@@ -630,7 +630,6 @@ defm int_amdgcn_global_atomic_fmin : noret_op;
 defm int_amdgcn_global_atomic_fmax : noret_op;
 defm int_amdgcn_global_atomic_csub : noret_op;
 defm int_amdgcn_flat_atomic_fadd : local_addr_space_atomic_op;
-defm int_amdgcn_ds_fadd_v2bf16 : noret_op;
 defm int_amdgcn_global_atomic_ordered_add_b64 : noret_op;
 defm int_amdgcn_flat_atomic_fmin_num : noret_op;
 defm int_amdgcn_flat_atomic_fmax_num : noret_op;

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -5403,8 +5403,6 @@ bool AMDGPULegalizerInfo::legalizeRsqClampIntrinsic(MachineInstr &MI,
 
 static unsigned getDSFPAtomicOpcode(Intrinsic::ID IID) {
   switch (IID) {
-  case Intrinsic::amdgcn_ds_fadd:
-    return AMDGPU::G_ATOMICRMW_FADD;
   case Intrinsic::amdgcn_ds_fmin:
     return AMDGPU::G_ATOMICRMW_FMIN;
   case Intrinsic::amdgcn_ds_fmax:
@@ -7333,7 +7331,6 @@ bool AMDGPULegalizerInfo::legalizeIntrinsic(LegalizerHelper &Helper,
     return legalizeBufferAtomic(MI, B, IntrID);
   case Intrinsic::amdgcn_rsq_clamp:
     return legalizeRsqClampIntrinsic(MI, MRI, B);
-  case Intrinsic::amdgcn_ds_fadd:
   case Intrinsic::amdgcn_ds_fmin:
   case Intrinsic::amdgcn_ds_fmax:
     return legalizeDSAtomicFPIntrinsic(Helper, MI, IntrID);

llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

@@ -4905,8 +4905,7 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
     case Intrinsic::amdgcn_global_load_tr_b128:
       return getDefaultMappingAllVGPR(MI);
     case Intrinsic::amdgcn_ds_ordered_add:
-    case Intrinsic::amdgcn_ds_ordered_swap:
-    case Intrinsic::amdgcn_ds_fadd_v2bf16: {
+    case Intrinsic::amdgcn_ds_ordered_swap: {
       unsigned DstSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
       OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, DstSize);
       unsigned M0Bank = getRegBankID(MI.getOperand(2).getReg(), MRI,

llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td

@@ -252,10 +252,8 @@ def : SourceOfDivergence<int_amdgcn_flat_atomic_fmin_num>;
 def : SourceOfDivergence<int_amdgcn_flat_atomic_fmax_num>;
 def : SourceOfDivergence<int_amdgcn_global_atomic_fadd_v2bf16>;
 def : SourceOfDivergence<int_amdgcn_flat_atomic_fadd_v2bf16>;
-def : SourceOfDivergence<int_amdgcn_ds_fadd>;
 def : SourceOfDivergence<int_amdgcn_ds_fmin>;
 def : SourceOfDivergence<int_amdgcn_ds_fmax>;
-def : SourceOfDivergence<int_amdgcn_ds_fadd_v2bf16>;
 def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_swap>;
 def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_add>;
 def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_sub>;

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

@@ -502,7 +502,6 @@ bool GCNTTIImpl::getTgtMemIntrinsic(IntrinsicInst *Inst,
   switch (Inst->getIntrinsicID()) {
   case Intrinsic::amdgcn_ds_ordered_add:
   case Intrinsic::amdgcn_ds_ordered_swap:
-  case Intrinsic::amdgcn_ds_fadd:
   case Intrinsic::amdgcn_ds_fmin:
   case Intrinsic::amdgcn_ds_fmax: {
     auto *Ordering = dyn_cast<ConstantInt>(Inst->getArgOperand(2));
@@ -1019,7 +1018,6 @@ bool GCNTTIImpl::isAlwaysUniform(const Value *V) const {
 bool GCNTTIImpl::collectFlatAddressOperands(SmallVectorImpl<int> &OpIndexes,
                                             Intrinsic::ID IID) const {
   switch (IID) {
-  case Intrinsic::amdgcn_ds_fadd:
   case Intrinsic::amdgcn_ds_fmin:
   case Intrinsic::amdgcn_ds_fmax:
   case Intrinsic::amdgcn_is_shared:
@@ -1041,7 +1039,6 @@ Value *GCNTTIImpl::rewriteIntrinsicWithAddressSpace(IntrinsicInst *II,
                                                     Value *NewV) const {
   auto IntrID = II->getIntrinsicID();
   switch (IntrID) {
-  case Intrinsic::amdgcn_ds_fadd:
   case Intrinsic::amdgcn_ds_fmin:
   case Intrinsic::amdgcn_ds_fmax: {
     const ConstantInt *IsVolatile = cast<ConstantInt>(II->getArgOperand(4));

llvm/lib/Target/AMDGPU/DSInstructions.td

@@ -1142,16 +1142,6 @@ def : DSAtomicRetPatIntrinsic<DS_ADD_F64, f64, int_amdgcn_flat_atomic_fadd_noret
 
 let SubtargetPredicate = HasAtomicDsPkAdd16Insts in {
 defm : DSAtomicRetNoRetPat_mc<DS_PK_ADD_RTN_F16, DS_PK_ADD_F16, v2f16, "atomic_load_fadd">;
-
-def : GCNPat <
-  (v2i16 (int_amdgcn_ds_fadd_v2bf16 i32:$ptr, v2i16:$src)),
-  (DS_PK_ADD_RTN_BF16 VGPR_32:$ptr, VGPR_32:$src, 0, 0)
->;
-let AddedComplexity = 1 in
-def : GCNPat <
-  (v2i16 (int_amdgcn_ds_fadd_v2bf16_noret i32:$ptr, v2i16:$src)),
-  (DS_PK_ADD_BF16 VGPR_32:$ptr, VGPR_32:$src, 0, 0)
->;
 } // End SubtargetPredicate = HasAtomicDsPkAdd16Insts
 
 let OtherPredicates = [HasGDS] in

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -1280,7 +1280,6 @@ bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
   switch (IntrID) {
   case Intrinsic::amdgcn_ds_ordered_add:
   case Intrinsic::amdgcn_ds_ordered_swap:
-  case Intrinsic::amdgcn_ds_fadd:
   case Intrinsic::amdgcn_ds_fmin:
   case Intrinsic::amdgcn_ds_fmax: {
     Info.opc = ISD::INTRINSIC_W_CHAIN;
@@ -1451,7 +1450,6 @@ bool SITargetLowering::getAddrModeArguments(IntrinsicInst *II,
   case Intrinsic::amdgcn_atomic_cond_sub_u32:
   case Intrinsic::amdgcn_ds_append:
   case Intrinsic::amdgcn_ds_consume:
-  case Intrinsic::amdgcn_ds_fadd:
   case Intrinsic::amdgcn_ds_fmax:
   case Intrinsic::amdgcn_ds_fmin:
   case Intrinsic::amdgcn_ds_ordered_add:
@@ -8700,19 +8698,6 @@ SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
                                    M->getVTList(), Ops, M->getMemoryVT(),
                                    M->getMemOperand());
   }
-  case Intrinsic::amdgcn_ds_fadd: {
-    MemSDNode *M = cast<MemSDNode>(Op);
-    unsigned Opc;
-    switch (IntrID) {
-    case Intrinsic::amdgcn_ds_fadd:
-      Opc = ISD::ATOMIC_LOAD_FADD;
-      break;
-    }
-
-    return DAG.getAtomic(Opc, SDLoc(Op), M->getMemoryVT(),
-                         M->getOperand(0), M->getOperand(2), M->getOperand(3),
-                         M->getMemOperand());
-  }
   case Intrinsic::amdgcn_ds_fmin:
   case Intrinsic::amdgcn_ds_fmax: {
     MemSDNode *M = cast<MemSDNode>(Op);