AMDGPU: Add subtarget feature for memory atomic fadd f64

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -788,6 +788,13 @@ def FeatureFlatAtomicFaddF32Inst
   "Has flat_atomic_add_f32 instruction"
 >;
 
+def FeatureFlatBufferGlobalAtomicFaddF64Inst
+  : SubtargetFeature<"flat-buffer-global-fadd-f64-inst",
+  "HasFlatBufferGlobalAtomicFaddF64Inst",
+  "true",
+  "Has flat, buffer, and global instructions for f64 atomic fadd"
+>;
+
 def FeatureMemoryAtomicFAddF32DenormalSupport
   : SubtargetFeature<"memory-atomic-fadd-f32-denormal-support",
   "HasMemoryAtomicFaddF32DenormalSupport",
@@ -1390,7 +1397,8 @@ def FeatureISAVersion9_0_A : FeatureSet<
      FeatureBackOffBarrier,
      FeatureKernargPreload,
      FeatureAtomicFMinFMaxF64GlobalInsts,
-     FeatureAtomicFMinFMaxF64FlatInsts
+     FeatureAtomicFMinFMaxF64FlatInsts,
+     FeatureFlatBufferGlobalAtomicFaddF64Inst
      ])>;
 
 def FeatureISAVersion9_0_C : FeatureSet<
@@ -1435,7 +1443,8 @@ def FeatureISAVersion9_4_Common : FeatureSet<
    FeatureAtomicFMinFMaxF64GlobalInsts,
    FeatureAtomicFMinFMaxF64FlatInsts,
    FeatureAgentScopeFineGrainedRemoteMemoryAtomics,
-   FeatureMemoryAtomicFAddF32DenormalSupport
+   FeatureMemoryAtomicFAddF32DenormalSupport,
+   FeatureFlatBufferGlobalAtomicFaddF64Inst
    ]>;
 
 def FeatureISAVersion9_4_0 : FeatureSet<
@@ -1932,11 +1941,9 @@ def isGFX12Plus :
 def HasFlatAddressSpace : Predicate<"Subtarget->hasFlatAddressSpace()">,
   AssemblerPredicate<(all_of FeatureFlatAddressSpace)>;
 
-
-def HasBufferFlatGlobalAtomicsF64 : // FIXME: Rename to show it's only for fadd
-  Predicate<"Subtarget->hasBufferFlatGlobalAtomicsF64()">,
-  // FIXME: This is too coarse, and working around using pseudo's predicates on real instruction.
-  AssemblerPredicate<(any_of FeatureGFX90AInsts, FeatureGFX10Insts, FeatureSouthernIslands, FeatureSeaIslands)>;
+def HasFlatBufferGlobalAtomicFaddF64Inst :
+  Predicate<"Subtarget->hasFlatBufferGlobalAtomicFaddF64Inst()">,
+  AssemblerPredicate<(any_of FeatureFlatBufferGlobalAtomicFaddF64Inst)>;
 
 def HasAtomicFMinFMaxF32GlobalInsts :
   Predicate<"Subtarget->hasAtomicFMinFMaxF32GlobalInsts()">,

llvm/lib/Target/AMDGPU/BUFInstructions.td

@@ -1312,14 +1312,16 @@ let SubtargetPredicate = isGFX90APlus in {
   }
 } // End SubtargetPredicate = isGFX90APlus
 
-let SubtargetPredicate = HasBufferFlatGlobalAtomicsF64 in {
+let SubtargetPredicate = HasFlatBufferGlobalAtomicFaddF64Inst in {
   defm BUFFER_ATOMIC_ADD_F64 : MUBUF_Pseudo_Atomics<"buffer_atomic_add_f64", VReg_64, f64>;
+} // End SubtargetPredicate = HasFlatBufferGlobalAtomicFaddF64Inst
 
+let SubtargetPredicate = HasAtomicFMinFMaxF64GlobalInsts in {
   // Note the names can be buffer_atomic_fmin_x2/buffer_atomic_fmax_x2
   // depending on some subtargets.
   defm BUFFER_ATOMIC_MIN_F64 : MUBUF_Pseudo_Atomics<"buffer_atomic_min_f64", VReg_64, f64>;
   defm BUFFER_ATOMIC_MAX_F64 : MUBUF_Pseudo_Atomics<"buffer_atomic_max_f64", VReg_64, f64>;
-} // End SubtargetPredicate = HasBufferFlatGlobalAtomicsF64
+}
 
 def BUFFER_INV : MUBUF_Invalidate<"buffer_inv"> {
   let SubtargetPredicate = isGFX940Plus;
@@ -1836,9 +1838,9 @@ let SubtargetPredicate = HasAtomicBufferGlobalPkAddF16Insts in {
   defm : SIBufferAtomicPat<"SIbuffer_atomic_fadd", v2f16, "BUFFER_ATOMIC_PK_ADD_F16", ["ret"]>;
 } // End SubtargetPredicate = HasAtomicBufferGlobalPkAddF16Insts
 
-let SubtargetPredicate = HasBufferFlatGlobalAtomicsF64 in {
+let SubtargetPredicate = HasFlatBufferGlobalAtomicFaddF64Inst in {
   defm : SIBufferAtomicPat<"SIbuffer_atomic_fadd", f64, "BUFFER_ATOMIC_ADD_F64">;
-} // End SubtargetPredicate = HasBufferFlatGlobalAtomicsF64
+} // End SubtargetPredicate = HasFlatBufferGlobalAtomicFaddF64Inst
 
 let SubtargetPredicate = HasAtomicFMinFMaxF64GlobalInsts in {
   defm : SIBufferAtomicPat<"SIbuffer_atomic_fmin", f64, "BUFFER_ATOMIC_MIN_F64">;

llvm/lib/Target/AMDGPU/FLATInstructions.td

@@ -772,10 +772,10 @@ defm GLOBAL_ATOMIC_MIN_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_min_f64",
 defm GLOBAL_ATOMIC_MAX_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_max_f64", VReg_64, f64>;
 }
 
-let SubtargetPredicate = HasBufferFlatGlobalAtomicsF64 in {
+let SubtargetPredicate = HasFlatBufferGlobalAtomicFaddF64Inst  in {
   defm FLAT_ATOMIC_ADD_F64   : FLAT_Atomic_Pseudo<"flat_atomic_add_f64", VReg_64, f64>;
   defm GLOBAL_ATOMIC_ADD_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_add_f64", VReg_64, f64>;
-} // End SubtargetPredicate = HasBufferFlatGlobalAtomicsF64
+} // End SubtargetPredicate = HasFlatBufferGlobalAtomicFaddF64Inst
 
 let SubtargetPredicate = HasAtomicFlatPkAdd16Insts in {
   defm FLAT_ATOMIC_PK_ADD_F16    : FLAT_Atomic_Pseudo<"flat_atomic_pk_add_f16",  VGPR_32, v2f16>;
@@ -1654,7 +1654,7 @@ defm : FlatAtomicIntrPat <"FLAT_ATOMIC_MIN_F64", "int_amdgcn_flat_atomic_fmin",
 defm : FlatAtomicIntrPat <"FLAT_ATOMIC_MAX_F64", "int_amdgcn_flat_atomic_fmax", f64>;
 }
 
-let OtherPredicates = [HasBufferFlatGlobalAtomicsF64] in {
+let OtherPredicates = [HasFlatBufferGlobalAtomicFaddF64Inst] in {
 defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_ADD_F64", "atomic_load_fadd_global", f64>;
 defm : GlobalFLATAtomicPatsWithAddrSpace<"GLOBAL_ATOMIC_ADD_F64", "int_amdgcn_flat_atomic_fadd", "global_addrspace", f64>;
 defm : GlobalFLATAtomicPatsWithAddrSpace<"GLOBAL_ATOMIC_ADD_F64", "int_amdgcn_global_atomic_fadd", "global_addrspace", f64>;

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -174,6 +174,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   bool HasAtomicGlobalPkAddBF16Inst = false;
   bool HasAtomicBufferPkAddBF16Inst = false;
   bool HasFlatAtomicFaddF32Inst = false;
+  bool HasFlatBufferGlobalAtomicFaddF64Inst = false;
   bool HasDefaultComponentZero = false;
   bool HasAgentScopeFineGrainedRemoteMemoryAtomics = false;
   bool HasDefaultComponentBroadcast = false;
@@ -660,9 +661,6 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
     return GFX10_BEncoding;
   }
 
-  // BUFFER/FLAT/GLOBAL_ATOMIC_ADD/MIN/MAX_F64
-  bool hasBufferFlatGlobalAtomicsF64() const { return hasGFX90AInsts(); }
-
   bool hasExportInsts() const {
     return !hasGFX940Insts();
   }
@@ -873,6 +871,12 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
 
   bool hasFlatAtomicFaddF32Inst() const { return HasFlatAtomicFaddF32Inst; }
 
+  /// \return true if the target has flat, global, and buffer atomic fadd for
+  /// double.
+  bool hasFlatBufferGlobalAtomicFaddF64Inst() const {
+    return HasFlatBufferGlobalAtomicFaddF64Inst;
+  }
+
   /// \return true if the target's flat, global, and buffer atomic fadd for
   /// float supports denormal handling.
   bool hasMemoryAtomicFaddF32DenormalSupport() const {

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -16229,7 +16229,7 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
       return AtomicExpansionKind::CmpXChg;
 
     // global and flat atomic fadd f64: gfx90a, gfx940.
-    if (Subtarget->hasGFX90AInsts() && Ty->isDoubleTy())
+    if (Subtarget->hasFlatBufferGlobalAtomicFaddF64Inst() && Ty->isDoubleTy())
       return ReportUnsafeHWInst(AtomicExpansionKind::None);
 
     if (AS != AMDGPUAS::FLAT_ADDRESS && Ty->isFloatTy()) {