AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory

Make this legal for gfx940 and gfx12

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -301,6 +301,9 @@ static const LLT V10S16 = LLT::fixed_vector(10, 16);
 static const LLT V12S16 = LLT::fixed_vector(12, 16);
 static const LLT V16S16 = LLT::fixed_vector(16, 16);
 
+static const LLT V2F16 = LLT::fixed_vector(2, LLT::float16());
+static const LLT V2BF16 = V2F16; // FIXME
+
 static const LLT V2S32 = LLT::fixed_vector(2, 32);
 static const LLT V3S32 = LLT::fixed_vector(3, 32);
 static const LLT V4S32 = LLT::fixed_vector(4, 32);
@@ -1638,7 +1641,7 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
     if (ST.hasLdsAtomicAddF64())
       Atomic.legalFor({{S64, LocalPtr}});
     if (ST.hasAtomicDsPkAdd16Insts())
-      Atomic.legalFor({{V2S16, LocalPtr}});
+      Atomic.legalFor({{V2F16, LocalPtr}, {V2BF16, LocalPtr}});
   }
   if (ST.hasAtomicFaddInsts())
     Atomic.legalFor({{S32, GlobalPtr}});

llvm/lib/Target/AMDGPU/DSInstructions.td

@@ -1082,6 +1082,12 @@ defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_U32, DS_MAX_U32, i32, "atomic_load_umax
 defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_F32, DS_MIN_F32, f32, "atomic_load_fmin">;
 defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_F32, DS_MAX_F32, f32, "atomic_load_fmax">;
 
+
+let SubtargetPredicate = HasAtomicDsPkAdd16Insts in {
+defm : DSAtomicRetNoRetPat_mc<DS_PK_ADD_RTN_F16, DS_PK_ADD_F16, v2f16, "atomic_load_fadd">;
+defm : DSAtomicRetNoRetPat_mc<DS_PK_ADD_RTN_BF16, DS_PK_ADD_BF16, v2bf16, "atomic_load_fadd">;
+}
+
 let SubtargetPredicate = isGFX6GFX7GFX8GFX9GFX10 in {
 defm : DSAtomicCmpXChgSwapped_mc<DS_CMPST_RTN_B32, DS_CMPST_B32, i32, "atomic_cmp_swap">;
 }

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -15931,6 +15931,16 @@ static OptimizationRemark emitAtomicRMWLegalRemark(const AtomicRMWInst *RMW) {
          << " operation at memory scope " << MemScope;
 }
 
+static bool isHalf2OrBFloat2(Type *Ty) {
+  if (FixedVectorType *VT = dyn_cast<FixedVectorType>(Ty)) {
+    Type *EltTy = VT->getElementType();
+    return VT->getNumElements() == 2 &&
+           (EltTy->isHalfTy() || EltTy->isBFloatTy());
+  }
+
+  return false;
+}
+
 TargetLowering::AtomicExpansionKind
 SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
   unsigned AS = RMW->getPointerAddressSpace();
@@ -15989,7 +15999,9 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
                                                  : AtomicExpansionKind::CmpXChg;
       }
 
-      // TODO: Handle v2f16/v2bf16 cases for gfx940
+      if (Subtarget->hasAtomicDsPkAdd16Insts() && isHalf2OrBFloat2(Ty))
+        return AtomicExpansionKind::None;
+
       return AtomicExpansionKind::CmpXChg;
     }
 

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

@@ -213,22 +213,8 @@ define <2 x half> @local_atomic_fadd_ret_v2f16_offset(ptr addrspace(3) %ptr, <2
 ; GFX940-LABEL: local_atomic_fadd_ret_v2f16_offset:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v2, v0 offset:65532
-; GFX940-NEXT:    s_mov_b64 s[0:1], 0
-; GFX940-NEXT:  .LBB15_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    ds_pk_add_rtn_f16 v0, v0, v1 offset:65532
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v3, v2
-; GFX940-NEXT:    v_pk_add_f16 v2, v3, v1
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    s_cbranch_execnz .LBB15_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
   %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst
@@ -239,21 +225,8 @@ define void @local_atomic_fadd_noret_v2f16_offset(ptr addrspace(3) %ptr, <2 x ha
 ; GFX940-LABEL: local_atomic_fadd_noret_v2f16_offset:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v2, v0 offset:65532
-; GFX940-NEXT:    s_mov_b64 s[0:1], 0
-; GFX940-NEXT:  .LBB16_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    ds_pk_add_f16 v0, v1 offset:65532
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_pk_add_f16 v3, v2, v1
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT:    v_mov_b32_e32 v2, v3
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    s_cbranch_execnz .LBB16_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
   %unused = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst

llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-atomicrmw.ll

@@ -52,36 +52,13 @@ define float @test_atomicrmw_fsub(ptr addrspace(3) %addr) {
 define <2 x half> @test_atomicrmw_fadd_vector(ptr addrspace(3) %addr) {
   ; CHECK-LABEL: name: test_atomicrmw_fadd_vector
   ; CHECK: bb.1 (%ir-block.0):
-  ; CHECK-NEXT:   successors: %bb.2(0x80000000)
   ; CHECK-NEXT:   liveins: $vgpr0
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0
   ; CHECK-NEXT:   [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3C00
   ; CHECK-NEXT:   [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16)
-  ; CHECK-NEXT:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
-  ; CHECK-NEXT:   [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p3) :: (load (<2 x s16>) from %ir.addr, addrspace 3)
-  ; CHECK-NEXT:   G_BR %bb.2
-  ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT: bb.2.atomicrmw.start:
-  ; CHECK-NEXT:   successors: %bb.3(0x40000000), %bb.2(0x40000000)
-  ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT:   [[PHI:%[0-9]+]]:_(s64) = G_PHI %19(s64), %bb.2, [[C1]](s64), %bb.1
-  ; CHECK-NEXT:   [[PHI1:%[0-9]+]]:_(<2 x s16>) = G_PHI [[LOAD]](<2 x s16>), %bb.1, %18(<2 x s16>), %bb.2
-  ; CHECK-NEXT:   [[FADD:%[0-9]+]]:_(<2 x s16>) = G_FADD [[PHI1]], [[BUILD_VECTOR]]
-  ; CHECK-NEXT:   [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[FADD]](<2 x s16>)
-  ; CHECK-NEXT:   [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[PHI1]](<2 x s16>)
-  ; CHECK-NEXT:   [[ATOMIC_CMPXCHG_WITH_SUCCESS:%[0-9]+]]:_(s32), [[ATOMIC_CMPXCHG_WITH_SUCCESS1:%[0-9]+]]:_(s1) = G_ATOMIC_CMPXCHG_WITH_SUCCESS [[COPY]](p3), [[BITCAST1]], [[BITCAST]] :: (load store seq_cst seq_cst (s32) on %ir.addr, addrspace 3)
-  ; CHECK-NEXT:   [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[ATOMIC_CMPXCHG_WITH_SUCCESS]](s32)
-  ; CHECK-NEXT:   [[INT:%[0-9]+]]:_(s64) = G_INTRINSIC intrinsic(@llvm.amdgcn.if.break), [[ATOMIC_CMPXCHG_WITH_SUCCESS1]](s1), [[PHI]](s64)
-  ; CHECK-NEXT:   [[INT1:%[0-9]+]]:_(s1) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.loop), [[INT]](s64)
-  ; CHECK-NEXT:   G_BRCOND [[INT1]](s1), %bb.3
-  ; CHECK-NEXT:   G_BR %bb.2
-  ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT: bb.3.atomicrmw.end:
-  ; CHECK-NEXT:   [[PHI2:%[0-9]+]]:_(<2 x s16>) = G_PHI [[BITCAST2]](<2 x s16>), %bb.2
-  ; CHECK-NEXT:   [[PHI3:%[0-9]+]]:_(s64) = G_PHI [[INT]](s64), %bb.2
-  ; CHECK-NEXT:   G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.end.cf), [[PHI3]](s64)
-  ; CHECK-NEXT:   $vgpr0 = COPY [[PHI2]](<2 x s16>)
+  ; CHECK-NEXT:   [[ATOMICRMW_FADD:%[0-9]+]]:_(<2 x s16>) = G_ATOMICRMW_FADD [[COPY]](p3), [[BUILD_VECTOR]] :: (load store seq_cst (<2 x s16>) on %ir.addr, addrspace 3)
+  ; CHECK-NEXT:   $vgpr0 = COPY [[ATOMICRMW_FADD]](<2 x s16>)
   ; CHECK-NEXT:   SI_RETURN implicit $vgpr0
   %oldval = atomicrmw fadd ptr addrspace(3) %addr, <2 x half> <half 1.0, half 1.0> seq_cst
   ret <2 x half> %oldval