clang/AMDGPU: Emit atomicrmw from flat_atomic_{f32|f64} builtins (#96874)

Author: arsenm

clang/lib/CodeGen/CGBuiltin.cpp

@@ -18931,10 +18931,8 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
   }
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fmin_f64:
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fmax_f64:
-  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64:
   case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmin_f64:
-  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmax_f64:
-  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32: {
+  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmax_f64: {
     Intrinsic::ID IID;
     llvm::Type *ArgTy = llvm::Type::getDoubleTy(getLLVMContext());
     switch (BuiltinID) {
@@ -18944,19 +18942,12 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
     case AMDGPU::BI__builtin_amdgcn_global_atomic_fmax_f64:
       IID = Intrinsic::amdgcn_global_atomic_fmax;
       break;
-    case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64:
-      IID = Intrinsic::amdgcn_flat_atomic_fadd;
-      break;
     case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmin_f64:
       IID = Intrinsic::amdgcn_flat_atomic_fmin;
       break;
     case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmax_f64:
       IID = Intrinsic::amdgcn_flat_atomic_fmax;
       break;
-    case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32:
-      ArgTy = llvm::Type::getFloatTy(getLLVMContext());
-      IID = Intrinsic::amdgcn_flat_atomic_fadd;
-      break;
     }
     llvm::Value *Addr = EmitScalarExpr(E->getArg(0));
     llvm::Value *Val = EmitScalarExpr(E->getArg(1));
@@ -19359,7 +19350,9 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32:
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64:
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16:
-  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2f16: {
+  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2f16:
+  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32:
+  case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64: {
     llvm::AtomicRMWInst::BinOp BinOp;
     switch (BuiltinID) {
     case AMDGPU::BI__builtin_amdgcn_atomic_inc32:
@@ -19379,6 +19372,8 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
     case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64:
     case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16:
     case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2f16:
+    case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32:
+    case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64:
       BinOp = llvm::AtomicRMWInst::FAdd;
       break;
     case AMDGPU::BI__builtin_amdgcn_ds_fminf:

clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx90a.cl

@@ -45,7 +45,8 @@ void test_global_max_f64(__global double *addr, double x){
 }
 
 // CHECK-LABEL: test_flat_add_local_f64
-// CHECK: call double @llvm.amdgcn.flat.atomic.fadd.f64.p3.f64(ptr addrspace(3) %{{.*}}, double %{{.*}})
+// CHECK: = atomicrmw fadd ptr addrspace(3) %{{.+}}, double %{{.+}} syncscope("agent") monotonic, align 8{{$}}
+
 // GFX90A-LABEL:  test_flat_add_local_f64$local
 // GFX90A:  ds_add_rtn_f64
 void test_flat_add_local_f64(__local double *addr, double x){
@@ -54,7 +55,8 @@ void test_flat_add_local_f64(__local double *addr, double x){
 }
 
 // CHECK-LABEL: test_flat_global_add_f64
-// CHECK: call double @llvm.amdgcn.flat.atomic.fadd.f64.p1.f64(ptr addrspace(1) %{{.*}}, double %{{.*}})
+// CHECK: = atomicrmw fadd ptr addrspace(1) {{.+}}, double %{{.+}} syncscope("agent") monotonic, align 8, !amdgpu.no.fine.grained.memory !{{[0-9]+$}}
+
 // GFX90A-LABEL:  test_flat_global_add_f64$local
 // GFX90A:  global_atomic_add_f64
 void test_flat_global_add_f64(__global double *addr, double x){

clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx940.cl

@@ -10,7 +10,8 @@ typedef half  __attribute__((ext_vector_type(2))) half2;
 typedef short __attribute__((ext_vector_type(2))) short2;
 
 // CHECK-LABEL: test_flat_add_f32
-// CHECK: call float @llvm.amdgcn.flat.atomic.fadd.f32.p0.f32(ptr %{{.*}}, float %{{.*}})
+// CHECK: [[RMW:%.+]] = atomicrmw fadd ptr %{{.+}}, float %{{.+}} syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !{{[0-9]+}}, !amdgpu.ignore.denormal.mode !{{[0-9]+$}}
+
 // GFX940-LABEL:  test_flat_add_f32
 // GFX940: flat_atomic_add_f32
 half2 test_flat_add_f32(__generic float *addr, float x) {