[AMDGPU] Remove support for no-return buffer atomic intrinsics. NFC. (#69326)

Thsi removes some of the machinery added by D85268, which was unused
since D87719 changed all buffer atomic intrinsics to return a value.

Author: jayfoad

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

@@ -1174,9 +1174,9 @@ class AMDGPUStructPtrBufferStore<LLVMType data_ty = llvm_any_ty> : DefaultAttrsI
 def int_amdgcn_struct_ptr_buffer_store_format : AMDGPUStructPtrBufferStore;
 def int_amdgcn_struct_ptr_buffer_store : AMDGPUStructPtrBufferStore;
 
-class AMDGPURawBufferAtomic<LLVMType data_ty = llvm_any_ty, bit NoRtn = false> : Intrinsic <
-  !if(NoRtn, [], [data_ty]),
-  [!if(NoRtn, data_ty, LLVMMatchType<0>),  // vdata(VGPR)
+class AMDGPURawBufferAtomic<LLVMType data_ty = llvm_any_ty> : Intrinsic <
+  [data_ty],
+  [LLVMMatchType<0>,  // vdata(VGPR)
    llvm_v4i32_ty,     // rsrc(SGPR)
    llvm_i32_ty,       // offset(VGPR/imm, included in bounds checking and swizzling)
    llvm_i32_ty,       // soffset(SGPR/imm, excluded from bounds checking and swizzling)
@@ -1208,9 +1208,9 @@ def int_amdgcn_raw_buffer_atomic_cmpswap : Intrinsic<
   [ImmArg<ArgIndex<5>>, IntrWillReturn, IntrNoCallback, IntrNoFree], "", [SDNPMemOperand]>,
   AMDGPURsrcIntrinsic<2, 0>;
 
-class AMDGPURawPtrBufferAtomic<LLVMType data_ty = llvm_any_ty, bit NoRtn = false> : Intrinsic <
-  !if(NoRtn, [], [data_ty]),
-  [!if(NoRtn, data_ty, LLVMMatchType<0>),  // vdata(VGPR)
+class AMDGPURawPtrBufferAtomic<LLVMType data_ty = llvm_any_ty> : Intrinsic <
+  [data_ty],
+  [LLVMMatchType<0>,            // vdata(VGPR)
    AMDGPUBufferRsrcTy,          // rsrc(SGPR)
    llvm_i32_ty,                 // offset(VGPR/imm, included in bounds checking and swizzling)
    llvm_i32_ty,                 // soffset(SGPR/imm, excluded from bounds checking and swizzling)
@@ -1249,9 +1249,9 @@ def int_amdgcn_raw_ptr_buffer_atomic_cmpswap : Intrinsic<
 def int_amdgcn_raw_buffer_atomic_fadd : AMDGPURawBufferAtomic<llvm_anyfloat_ty>;
 def int_amdgcn_raw_ptr_buffer_atomic_fadd : AMDGPURawPtrBufferAtomic<llvm_anyfloat_ty>;
 
-class AMDGPUStructBufferAtomic<LLVMType data_ty = llvm_any_ty, bit NoRtn = false> : Intrinsic <
-  !if(NoRtn, [], [data_ty]),
-  [!if(NoRtn, data_ty, LLVMMatchType<0>),  // vdata(VGPR)
+class AMDGPUStructBufferAtomic<LLVMType data_ty = llvm_any_ty> : Intrinsic <
+  [data_ty],
+  [LLVMMatchType<0>,  // vdata(VGPR)
    llvm_v4i32_ty,     // rsrc(SGPR)
    llvm_i32_ty,       // vindex(VGPR)
    llvm_i32_ty,       // offset(VGPR/imm, included in bounds checking and swizzling)
@@ -1283,9 +1283,9 @@ def int_amdgcn_struct_buffer_atomic_cmpswap : Intrinsic<
   [ImmArg<ArgIndex<6>>, IntrWillReturn, IntrNoCallback, IntrNoFree], "", [SDNPMemOperand]>,
   AMDGPURsrcIntrinsic<2, 0>;
 
-class AMDGPUStructPtrBufferAtomic<LLVMType data_ty = llvm_any_ty, bit NoRtn = false> : Intrinsic <
-  !if(NoRtn, [], [data_ty]),
-  [!if(NoRtn, data_ty, LLVMMatchType<0>),  // vdata(VGPR)
+class AMDGPUStructPtrBufferAtomic<LLVMType data_ty = llvm_any_ty> : Intrinsic <
+  [data_ty],
+  [LLVMMatchType<0>,            // vdata(VGPR)
    AMDGPUBufferRsrcTy,          // rsrc(SGPR)
    llvm_i32_ty,                 // vindex(VGPR)
    llvm_i32_ty,                 // offset(VGPR/imm, included in bounds checking and swizzling)

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -5879,31 +5879,23 @@ bool AMDGPULegalizerInfo::legalizeBufferAtomic(MachineInstr &MI,
       IID == Intrinsic::amdgcn_struct_buffer_atomic_cmpswap ||
       IID == Intrinsic::amdgcn_raw_ptr_buffer_atomic_cmpswap ||
       IID == Intrinsic::amdgcn_struct_ptr_buffer_atomic_cmpswap;
-  const bool HasReturn = MI.getNumExplicitDefs() != 0;
-
-  Register Dst;
-
-  int OpOffset = 0;
-  if (HasReturn) {
-    // A few FP atomics do not support return values.
-    Dst = MI.getOperand(0).getReg();
-  } else {
-    OpOffset = -1;
-  }
 
+  Register Dst = MI.getOperand(0).getReg();
   // Since we don't have 128-bit atomics, we don't need to handle the case of
   // p8 argmunents to the atomic itself
-  Register VData = MI.getOperand(2 + OpOffset).getReg();
+  Register VData = MI.getOperand(2).getReg();
+
   Register CmpVal;
+  int OpOffset = 0;
 
   if (IsCmpSwap) {
-    CmpVal = MI.getOperand(3 + OpOffset).getReg();
+    CmpVal = MI.getOperand(3).getReg();
     ++OpOffset;
   }
 
   castBufferRsrcArgToV4I32(MI, B, 3 + OpOffset);
   Register RSrc = MI.getOperand(3 + OpOffset).getReg();
-  const unsigned NumVIndexOps = (IsCmpSwap ? 8 : 7) + HasReturn;
+  const unsigned NumVIndexOps = IsCmpSwap ? 9 : 8;
 
   // The struct intrinsic variants add one additional operand over raw.
   const bool HasVIndex = MI.getNumOperands() == NumVIndexOps;
@@ -5924,12 +5916,9 @@ bool AMDGPULegalizerInfo::legalizeBufferAtomic(MachineInstr &MI,
   unsigned ImmOffset;
   std::tie(VOffset, ImmOffset) = splitBufferOffsets(B, VOffset);
 
-  auto MIB = B.buildInstr(getBufferAtomicPseudo(IID));
-
-  if (HasReturn)
-    MIB.addDef(Dst);
-
-  MIB.addUse(VData); // vdata
+  auto MIB = B.buildInstr(getBufferAtomicPseudo(IID))
+      .addDef(Dst)
+      .addUse(VData); // vdata
 
   if (IsCmpSwap)
     MIB.addReg(CmpVal);

llvm/lib/Target/AMDGPU/SIInstructions.td

@@ -3572,8 +3572,8 @@ def G_AMDGPU_ATOMIC_FMIN : G_ATOMICRMW_OP;
 def G_AMDGPU_ATOMIC_FMAX : G_ATOMICRMW_OP;
 }
 
-class BufferAtomicGenericInstruction<bit NoRtn = 0> : AMDGPUGenericInstruction {
-  let OutOperandList = !if(NoRtn, (outs), (outs type0:$dst));
+class BufferAtomicGenericInstruction : AMDGPUGenericInstruction {
+  let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins type0:$vdata, type1:$rsrc, type2:$vindex, type2:$voffset,
                            type2:$soffset, untyped_imm_0:$offset,
                            untyped_imm_0:$cachepolicy, untyped_imm_0:$idxen);