AMDGPU: Remove flat/global atomic fadd v2bf16 intrinsics (llvm#97050)

These are now fully covered by atomicrmw.

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

@@ -2996,10 +2996,6 @@ multiclass AMDGPUMFp8SmfmacIntrinsic<LLVMType DestTy> {
     def NAME#"_"#kind : AMDGPUMFp8SmfmacIntrinsic<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>;
-
 defset list<Intrinsic> AMDGPUMFMAIntrinsics940 = {
 def int_amdgcn_mfma_i32_16x16x32_i8     : AMDGPUMfmaIntrinsic<llvm_v4i32_ty,  llvm_i64_ty>;
 def int_amdgcn_mfma_i32_32x32x16_i8     : AMDGPUMfmaIntrinsic<llvm_v16i32_ty, llvm_i64_ty>;

llvm/lib/IR/AutoUpgrade.cpp

@@ -1034,7 +1034,9 @@ static bool upgradeIntrinsicFunction1(Function *F, Function *&NewFn,
       }
 
       if (Name.starts_with("ds.fadd") || Name.starts_with("ds.fmin") ||
-          Name.starts_with("ds.fmax")) {
+          Name.starts_with("ds.fmax") ||
+          Name.starts_with("global.atomic.fadd.v2bf16") ||
+          Name.starts_with("flat.atomic.fadd.v2bf16")) {
         // Replaced with atomicrmw fadd/fmin/fmax, so there's no new
         // declaration.
         NewFn = nullptr;
@@ -4042,7 +4044,9 @@ static Value *upgradeAMDGCNIntrinsicCall(StringRef Name, CallBase *CI,
           .StartsWith("ds.fmin", AtomicRMWInst::FMin)
           .StartsWith("ds.fmax", AtomicRMWInst::FMax)
           .StartsWith("atomic.inc.", AtomicRMWInst::UIncWrap)
-          .StartsWith("atomic.dec.", AtomicRMWInst::UDecWrap);
+          .StartsWith("atomic.dec.", AtomicRMWInst::UDecWrap)
+          .StartsWith("global.atomic.fadd", AtomicRMWInst::FAdd)
+          .StartsWith("flat.atomic.fadd", AtomicRMWInst::FAdd);
 
   unsigned NumOperands = CI->getNumOperands();
   if (NumOperands < 3) // Malformed bitcode.
@@ -4097,8 +4101,10 @@ static Value *upgradeAMDGCNIntrinsicCall(StringRef Name, CallBase *CI,
       Builder.CreateAtomicRMW(RMWOp, Ptr, Val, std::nullopt, Order, SSID);
 
   if (PtrTy->getAddressSpace() != 3) {
-    RMW->setMetadata("amdgpu.no.fine.grained.memory",
-                     MDNode::get(F->getContext(), {}));
+    MDNode *EmptyMD = MDNode::get(F->getContext(), {});
+    RMW->setMetadata("amdgpu.no.fine.grained.memory", EmptyMD);
+    if (RMWOp == AtomicRMWInst::FAdd && RetTy->isFloatTy())
+      RMW->setMetadata("amdgpu.ignore.denormal.mode", EmptyMD);
   }
 
   if (IsVolatile)

llvm/lib/Target/AMDGPU/AMDGPUInstructions.td

@@ -620,12 +620,10 @@ multiclass local_addr_space_atomic_op {
 
 defm int_amdgcn_flat_atomic_fadd : noret_op;
 defm int_amdgcn_flat_atomic_fadd : flat_addr_space_atomic_op;
-defm int_amdgcn_flat_atomic_fadd_v2bf16 : noret_op;
 defm int_amdgcn_flat_atomic_fmin : noret_op;
 defm int_amdgcn_flat_atomic_fmax : noret_op;
 defm int_amdgcn_global_atomic_fadd : global_addr_space_atomic_op;
 defm int_amdgcn_flat_atomic_fadd : global_addr_space_atomic_op;
-defm int_amdgcn_global_atomic_fadd_v2bf16 : noret_op;
 defm int_amdgcn_global_atomic_fmin : noret_op;
 defm int_amdgcn_global_atomic_fmax : noret_op;
 defm int_amdgcn_global_atomic_csub : noret_op;

llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

@@ -4907,8 +4907,6 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
     case Intrinsic::amdgcn_flat_atomic_fmax:
     case Intrinsic::amdgcn_flat_atomic_fmin_num:
     case Intrinsic::amdgcn_flat_atomic_fmax_num:
-    case Intrinsic::amdgcn_global_atomic_fadd_v2bf16:
-    case Intrinsic::amdgcn_flat_atomic_fadd_v2bf16:
     case Intrinsic::amdgcn_atomic_cond_sub_u32:
     case Intrinsic::amdgcn_global_atomic_ordered_add_b64:
     case Intrinsic::amdgcn_global_load_tr_b64:

llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td

@@ -250,8 +250,6 @@ def : SourceOfDivergence<int_amdgcn_flat_atomic_fmin>;
 def : SourceOfDivergence<int_amdgcn_flat_atomic_fmax>;
 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_raw_buffer_atomic_swap>;
 def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_add>;
 def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_sub>;

llvm/lib/Target/AMDGPU/FLATInstructions.td

@@ -1674,13 +1674,11 @@ defm : FlatAtomicIntrPat <"FLAT_ATOMIC_ADD_F32", "int_amdgcn_flat_atomic_fadd",
 
 let OtherPredicates = [HasAtomicFlatPkAdd16Insts] in {
 defm : FlatAtomicIntrPat <"FLAT_ATOMIC_PK_ADD_F16", "int_amdgcn_flat_atomic_fadd", v2f16>;
-defm : FlatAtomicIntrPat <"FLAT_ATOMIC_PK_ADD_BF16", "int_amdgcn_flat_atomic_fadd_v2bf16", v2i16>;
 defm : FlatAtomicPat <"FLAT_ATOMIC_PK_ADD_F16", "atomic_load_fadd_flat", v2f16>;
 defm : FlatAtomicPat <"FLAT_ATOMIC_PK_ADD_BF16", "atomic_load_fadd_flat", v2bf16>;
 }
 
 let OtherPredicates = [HasAtomicGlobalPkAddBF16Inst] in
-defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_PK_ADD_BF16", "int_amdgcn_global_atomic_fadd_v2bf16", v2i16>;
 defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_PK_ADD_BF16", "atomic_load_fadd_global", v2bf16>;
 } // End OtherPredicates = [HasFlatGlobalInsts], AddedComplexity = 10
 

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -1362,9 +1362,7 @@ bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
   case Intrinsic::amdgcn_flat_atomic_fmax:
   case Intrinsic::amdgcn_flat_atomic_fmin_num:
   case Intrinsic::amdgcn_flat_atomic_fmax_num:
-  case Intrinsic::amdgcn_global_atomic_fadd_v2bf16:
-  case Intrinsic::amdgcn_atomic_cond_sub_u32:
-  case Intrinsic::amdgcn_flat_atomic_fadd_v2bf16: {
+  case Intrinsic::amdgcn_atomic_cond_sub_u32: {
     Info.opc = ISD::INTRINSIC_W_CHAIN;
     Info.memVT = MVT::getVT(CI.getType());
     Info.ptrVal = CI.getOperand(0);
@@ -1467,14 +1465,12 @@ bool SITargetLowering::getAddrModeArguments(IntrinsicInst *II,
   case Intrinsic::amdgcn_ds_ordered_add:
   case Intrinsic::amdgcn_ds_ordered_swap:
   case Intrinsic::amdgcn_flat_atomic_fadd:
-  case Intrinsic::amdgcn_flat_atomic_fadd_v2bf16:
   case Intrinsic::amdgcn_flat_atomic_fmax:
   case Intrinsic::amdgcn_flat_atomic_fmax_num:
   case Intrinsic::amdgcn_flat_atomic_fmin:
   case Intrinsic::amdgcn_flat_atomic_fmin_num:
   case Intrinsic::amdgcn_global_atomic_csub:
   case Intrinsic::amdgcn_global_atomic_fadd:
-  case Intrinsic::amdgcn_global_atomic_fadd_v2bf16:
   case Intrinsic::amdgcn_global_atomic_fmax:
   case Intrinsic::amdgcn_global_atomic_fmax_num:
   case Intrinsic::amdgcn_global_atomic_fmin:

llvm/test/Bitcode/amdgcn-atomic.ll

@@ -300,4 +300,26 @@ define float @upgrade_amdgcn_ds_fmax_f32_no_suffix(ptr addrspace(3) %ptr, float
   ret float %result0
 }
 
+declare <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0(ptr, <2 x i16>)
+
+define <2 x i16> @upgrade_amdgcn_flat_atomic_fadd_v2bf16_p0(ptr %ptr, <2 x i16> %data) {
+  ; CHECK: [[BC0:%.+]] = bitcast <2 x i16> %data to <2 x bfloat>
+  ; CHECK-NEXT: [[ATOMIC:%.+]] = atomicrmw fadd ptr %ptr, <2 x bfloat> [[BC0]] syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory !{{[0-9]+$}}
+  ; CHECK-NEXT: [[BC1:%.+]] = bitcast <2 x bfloat> [[ATOMIC]] to <2 x i16>
+  ; CHECK-NEXT: ret <2 x i16> [[BC1]]
+  %result = call <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0(ptr %ptr, <2 x i16> %data)
+  ret <2 x i16> %result
+}
+
+declare <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1(ptr addrspace(1), <2 x i16>)
+
+define <2 x i16> @upgrade_amdgcn_global_atomic_fadd_v2bf16_p1(ptr addrspace(1) %ptr, <2 x i16> %data) {
+  ; CHECK: [[BC0:%.+]] = bitcast <2 x i16> %data to <2 x bfloat>
+  ; CHECK-NEXT: [[ATOMIC:%.+]] = atomicrmw fadd ptr addrspace(1) %ptr, <2 x bfloat> [[BC0]] syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory !{{[0-9]+$}}
+  ; CHECK-NEXT: [[BC1:%.+]] = bitcast <2 x bfloat> [[ATOMIC]] to <2 x i16>
+  ; CHECK-NEXT: ret <2 x i16> [[BC1]]
+  %result = call <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1(ptr addrspace(1) %ptr, <2 x i16> %data)
+  ret <2 x i16> %result
+}
+
 attributes #0 = { argmemonly nounwind willreturn }

llvm/test/CodeGen/AMDGPU/GlobalISel/fp-atomics-gfx940.ll

@@ -4,12 +4,6 @@
 declare float @llvm.amdgcn.flat.atomic.fadd.f32.p0.f32(ptr %ptr, float %data)
 declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p0.v2f16(ptr %ptr, <2 x half> %data)
 
-; bf16 atomics use v2i16 argument since there is no bf16 data type in the llvm.
-declare <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0(ptr %ptr, <2 x i16> %data)
-declare <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1(ptr addrspace(1) %ptr, <2 x i16> %data)
-declare <2 x half> @llvm.amdgcn.ds.fadd.v2f16(ptr addrspace(3) %ptr, <2 x half> %data, i32, i32, i1)
-declare <2 x i16> @llvm.amdgcn.ds.fadd.v2bf16(ptr addrspace(3) %ptr, <2 x i16> %data)
-
 define amdgpu_kernel void @flat_atomic_fadd_f32_noret(ptr %ptr, float %data) {
 ; GFX940-LABEL: flat_atomic_fadd_f32_noret:
 ; GFX940:       ; %bb.0:
@@ -106,106 +100,6 @@ define <2 x half> @flat_atomic_fadd_v2f16_rtn(ptr %ptr, <2 x half> %data) {
   ret <2 x half> %ret
 }
 
-define amdgpu_kernel void @flat_atomic_fadd_v2bf16_noret(ptr %ptr, <2 x i16> %data) {
-; GFX940-LABEL: flat_atomic_fadd_v2bf16_noret:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
-; GFX940-NEXT:    s_load_dword s4, s[2:3], 0x2c
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[0:1]
-; GFX940-NEXT:    v_mov_b32_e32 v2, s4
-; GFX940-NEXT:    flat_atomic_pk_add_bf16 v[0:1], v2
-; GFX940-NEXT:    s_endpgm
-  %ret = call <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0(ptr %ptr, <2 x i16> %data)
-  ret void
-}
-
-define <2 x i16> @flat_atomic_fadd_v2bf16_rtn(ptr %ptr, <2 x i16> %data) {
-; GFX940-LABEL: flat_atomic_fadd_v2bf16_rtn:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    flat_atomic_pk_add_bf16 v0, v[0:1], v2 sc0
-; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    s_setpc_b64 s[30:31]
-  %ret = call <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0(ptr %ptr, <2 x i16> %data)
-  ret <2 x i16> %ret
-}
-
-define amdgpu_kernel void @global_atomic_fadd_v2bf16_noret(ptr addrspace(1) %ptr, <2 x i16> %data) {
-; GFX940-LABEL: global_atomic_fadd_v2bf16_noret:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_load_dword s4, s[2:3], 0x2c
-; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
-; GFX940-NEXT:    v_mov_b32_e32 v1, 0
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v0, s4
-; GFX940-NEXT:    global_atomic_pk_add_bf16 v1, v0, s[0:1]
-; GFX940-NEXT:    s_endpgm
-  %ret = call <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1(ptr addrspace(1) %ptr, <2 x i16> %data)
-  ret void
-}
-
-define <2 x i16> @global_atomic_fadd_v2bf16_rtn(ptr addrspace(1) %ptr, <2 x i16> %data) {
-; GFX940-LABEL: global_atomic_fadd_v2bf16_rtn:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    global_atomic_pk_add_bf16 v0, v[0:1], v2, off sc0
-; GFX940-NEXT:    s_waitcnt vmcnt(0)
-; GFX940-NEXT:    s_setpc_b64 s[30:31]
-  %ret = call <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1(ptr addrspace(1) %ptr, <2 x i16> %data)
-  ret <2 x i16> %ret
-}
-
-define amdgpu_kernel void @local_atomic_fadd_v2f16_noret(ptr addrspace(3) %ptr, <2 x half> %data) {
-; GFX940-LABEL: local_atomic_fadd_v2f16_noret:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v0, s0
-; GFX940-NEXT:    v_mov_b32_e32 v1, s1
-; GFX940-NEXT:    ds_pk_add_f16 v0, v1
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    s_endpgm
-  %ret = call <2 x half> @llvm.amdgcn.ds.fadd.v2f16(ptr addrspace(3) %ptr, <2 x half> %data, i32 0, i32 0, i1 0)
-  ret void
-}
-
-define <2 x half> @local_atomic_fadd_v2f16_rtn(ptr addrspace(3) %ptr, <2 x half> %data) {
-; GFX940-LABEL: local_atomic_fadd_v2f16_rtn:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_pk_add_rtn_f16 v0, v0, v1
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    s_setpc_b64 s[30:31]
-  %ret = call <2 x half> @llvm.amdgcn.ds.fadd.v2f16(ptr addrspace(3) %ptr, <2 x half> %data, i32 0, i32 0, i1 0)
-  ret <2 x half> %ret
-}
-
-define amdgpu_kernel void @local_atomic_fadd_v2bf16_noret(ptr addrspace(3) %ptr, <2 x i16> %data) {
-; GFX940-LABEL: local_atomic_fadd_v2bf16_noret:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v0, s0
-; GFX940-NEXT:    v_mov_b32_e32 v1, s1
-; GFX940-NEXT:    ds_pk_add_f16 v0, v1
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    s_endpgm
-  %ret = call <2 x i16> @llvm.amdgcn.ds.fadd.v2bf16(ptr addrspace(3) %ptr, <2 x i16> %data)
-  ret void
-}
-
-define <2 x i16> @local_atomic_fadd_v2bf16_rtn(ptr addrspace(3) %ptr, <2 x i16> %data) {
-; GFX940-LABEL: local_atomic_fadd_v2bf16_rtn:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_pk_add_rtn_f16 v0, v0, v1
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    s_setpc_b64 s[30:31]
-  %ret = call <2 x i16> @llvm.amdgcn.ds.fadd.v2bf16(ptr addrspace(3) %ptr, <2 x i16> %data)
-  ret <2 x i16> %ret
-}
-
 define <2 x half> @local_atomic_fadd_ret_v2f16_offset(ptr addrspace(3) %ptr, <2 x half> %val) {
 ; GFX940-LABEL: local_atomic_fadd_ret_v2f16_offset:
 ; GFX940:       ; %bb.0: