[NVPTX] Implement __nvvm_atom_add_gen_d builtin.

[jlebar]

Summary

This just seems to have been an oversight. We already supported the f64
atomic add with an explicit scope (e.g. "cta"), but not the scopeless
version.

[jlebar]

Fix tests.

clang/include/clang/Basic/BuiltinsNVPTX.def

 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_ll, "LLiLLiD*LLi", "n", "satom")
 BUILTIN(__nvvm_atom_add_g_f, "ffD*1f", "n")
 BUILTIN(__nvvm_atom_add_s_f, "ffD*3f", "n")
 BUILTIN(__nvvm_atom_add_gen_f, "ffD*f", "n")
 TARGET_BUILTIN(__nvvm_atom_cta_add_gen_f, "ffD*f", "n", "satom")
 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_f, "ffD*f", "n", "satom")
 BUILTIN(__nvvm_atom_add_g_d, "ddD*1d", "n")
 BUILTIN(__nvvm_atom_add_s_d, "ddD*3d", "n")
-BUILTIN(__nvvm_atom_add_gen_d, "ddD*d", "n")
+TARGET_BUILTIN(__nvvm_atom_add_gen_d, "ddD*d", "n", "satom")
 TARGET_BUILTIN(__nvvm_atom_cta_add_gen_d, "ddD*d", "n", "satom")
 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_d, "ddD*d", "n", "satom")
 
 BUILTIN(__nvvm_atom_sub_g_i, "iiD*1i", "n")
 BUILTIN(__nvvm_atom_sub_s_i, "iiD*3i", "n")
 BUILTIN(__nvvm_atom_sub_gen_i, "iiD*i", "n")
 BUILTIN(__nvvm_atom_sub_g_l, "LiLiD*1Li", "n")
 BUILTIN(__nvvm_atom_sub_s_l, "LiLiD*3Li", "n")

clang/lib/CodeGen/CGBuiltin.cpp

case NVPTX::BI__nvvm_atom_add_gen_f: {
     Value *Val = EmitScalarExpr(E->getArg(1));
     // atomicrmw only deals with integer arguments so we need to use
     // LLVM's nvvm_atomic_load_add_f32 intrinsic for that.
     Value *FnALAF32 =
         CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType());
     return Builder.CreateCall(FnALAF32, {Ptr, Val});
   }
 
+  case NVPTX::BI__nvvm_atom_add_gen_d: {
+    Value *Ptr = EmitScalarExpr(E->getArg(0));
+    Value *Val = EmitScalarExpr(E->getArg(1));
+    // atomicrmw only deals with integer arguments, so we need to use
+    // LLVM's nvvm_atomic_load_add_f64 intrinsic.
+    Value *FnALAF64 =
+        CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f64, Ptr->getType());
+    return Builder.CreateCall(FnALAF64, {Ptr, Val});
+  }
+
   case NVPTX::BI__nvvm_atom_inc_gen_ui: {
     Value *Ptr = EmitScalarExpr(E->getArg(0));
     Value *Val = EmitScalarExpr(E->getArg(1));
     Value *FnALI32 =
         CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType());
     return Builder.CreateCall(FnALI32, {Ptr, Val});
   }
 

clang/test/CodeGen/builtins-nvptx-ptx50.cu

+// RUN: %clang_cc1 -triple nvptx64-unknown-unknown -target-cpu sm_60 \
+// RUN:            -fcuda-is-device -S -emit-llvm -o - -x cuda %s \
+// RUN:   | FileCheck -check-prefix=CHECK %s
+//
+// RUN: %clang_cc1 -triple nvptx-unknown-unknown -target-cpu sm_50 \
+// RUN:   -fcuda-is-device -S -o /dev/null -x cuda -verify %s
+
+#define __device__ __attribute__((device))
+#define __global__ __attribute__((global))
+#define __shared__ __attribute__((shared))
+#define __constant__ __attribute__((constant))
+
+// We have to keep all builtins that depend on particular target feature in the
+// same function, because the codegen will stop after the very first function
+// that encounters an error, so -verify will not be able to find errors in
+// subsequent functions.
+
+// CHECK-LABEL: test_fn
+__device__ void test_fn(double d, double* double_ptr) {
+  // CHECK: call double @llvm.nvvm.atomic.load.add.f64.p0f64
+  // expected-error@+1 {{'__nvvm_atom_add_gen_d' needs target feature satom}}
+  __nvvm_atom_add_gen_d(double_ptr, d);
+}

llvm/include/llvm/IR/IntrinsicsNVVM.td

def int_nvvm_bitcast_i2f : GCCBuiltin<"__nvvm_bitcast_i2f">,
       Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
 
   def int_nvvm_bitcast_ll2d : GCCBuiltin<"__nvvm_bitcast_ll2d">,
       Intrinsic<[llvm_double_ty], [llvm_i64_ty], [IntrNoMem]>;
   def int_nvvm_bitcast_d2ll : GCCBuiltin<"__nvvm_bitcast_d2ll">,
       Intrinsic<[llvm_i64_ty], [llvm_double_ty], [IntrNoMem]>;
 
 
-// Atomic not available as an llvm intrinsic.
+// Atomics not available as llvm intrinsics.
   def int_nvvm_atomic_load_add_f32 : Intrinsic<[llvm_float_ty],
           [LLVMAnyPointerType<llvm_float_ty>, llvm_float_ty],
                                       [IntrArgMemOnly, NoCapture<0>]>;
+  // Atomic add of f64 requires sm_60.
+  def int_nvvm_atomic_load_add_f64 : Intrinsic<[llvm_double_ty],
+          [LLVMAnyPointerType<llvm_double_ty>, llvm_double_ty],
+                                      [IntrArgMemOnly, NoCapture<0>]>;
+
   def int_nvvm_atomic_load_inc_32 : Intrinsic<[llvm_i32_ty],
           [LLVMAnyPointerType<llvm_i32_ty>, llvm_i32_ty],
                                       [IntrArgMemOnly, NoCapture<0>]>;
   def int_nvvm_atomic_load_dec_32 : Intrinsic<[llvm_i32_ty],
           [LLVMAnyPointerType<llvm_i32_ty>, llvm_i32_ty],
                                       [IntrArgMemOnly, NoCapture<0>]>;
 
   class SCOPED_ATOMIC2_impl<LLVMType elty>

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

case Intrinsic::nvvm_wmma_store_d_f32_row_global_stride: {
     Info.vol = false;
     Info.readMem = false;
     Info.writeMem = true;
     Info.align = 16;
     return true;
   }
 
   case Intrinsic::nvvm_atomic_load_add_f32:
+  case Intrinsic::nvvm_atomic_load_add_f64:
   case Intrinsic::nvvm_atomic_load_inc_32:
   case Intrinsic::nvvm_atomic_load_dec_32:
 
   case Intrinsic::nvvm_atomic_add_gen_f_cta:
   case Intrinsic::nvvm_atomic_add_gen_f_sys:
   case Intrinsic::nvvm_atomic_add_gen_i_cta:
   case Intrinsic::nvvm_atomic_add_gen_i_sys:
   case Intrinsic::nvvm_atomic_and_gen_i_cta:

llvm/lib/Target/NVPTX/NVPTXIntrinsics.td

 def atomic_load_add_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
   (atomic_load_add_64 node:$a, node:$b)>;
 def atomic_load_add_f32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
   (int_nvvm_atomic_load_add_f32 node:$a, node:$b)>;
 def atomic_load_add_f32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
   (int_nvvm_atomic_load_add_f32 node:$a, node:$b)>;
 def atomic_load_add_f32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
   (int_nvvm_atomic_load_add_f32 node:$a, node:$b)>;
+def atomic_load_add_f64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
+  (int_nvvm_atomic_load_add_f64 node:$a, node:$b)>;
+def atomic_load_add_f64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
+  (int_nvvm_atomic_load_add_f64 node:$a, node:$b)>;
+def atomic_load_add_f64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
+  (int_nvvm_atomic_load_add_f64 node:$a, node:$b)>;
 
 defm INT_PTX_ATOM_ADD_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".u32", ".add",
   atomic_load_add_32_g, i32imm, imm, hasAtomRedG32>;
 defm INT_PTX_ATOM_ADD_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".u32", ".add",
   atomic_load_add_32_s, i32imm, imm, hasAtomRedS32>;
 defm INT_PTX_ATOM_ADD_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".u32", ".add",
   atomic_load_add_32_gen, i32imm, imm, hasAtomRedGen32>;
 defm INT_PTX_ATOM_ADD_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".u32",

 
 defm INT_PTX_ATOM_ADD_G_F32 : F_ATOMIC_2<Float32Regs, ".global", ".f32", ".add",
   atomic_load_add_f32_g, f32imm, fpimm, hasAtomAddF32>;
 defm INT_PTX_ATOM_ADD_S_F32 : F_ATOMIC_2<Float32Regs, ".shared", ".f32", ".add",
   atomic_load_add_f32_s, f32imm, fpimm, hasAtomAddF32>;
 defm INT_PTX_ATOM_ADD_GEN_F32 : F_ATOMIC_2<Float32Regs, "", ".f32", ".add",
   atomic_load_add_f32_gen, f32imm, fpimm, hasAtomAddF32>;
 
+defm INT_PTX_ATOM_ADD_G_F64 : F_ATOMIC_2<Float64Regs, ".global", ".f64", ".add",
+  atomic_load_add_f64_g, f64imm, fpimm, hasAtomAddF64>;
+defm INT_PTX_ATOM_ADD_S_F64 : F_ATOMIC_2<Float64Regs, ".shared", ".f64", ".add",
+  atomic_load_add_f64_s, f64imm, fpimm, hasAtomAddF64>;
+defm INT_PTX_ATOM_ADD_GEN_F64 : F_ATOMIC_2<Float64Regs, "", ".f64", ".add",
+  atomic_load_add_f64_gen, f64imm, fpimm, hasAtomAddF64>;
+
 // atom_sub
 
 def atomic_load_sub_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
   (atomic_load_sub_32 node:$a, node:$b)>;
 def atomic_load_sub_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
   (atomic_load_sub_32 node:$a, node:$b)>;
 def atomic_load_sub_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
   (atomic_load_sub_32 node:$a, node:$b)>;

llvm/test/CodeGen/NVPTX/atomics-sm60.ll

+; RUN: llc < %s -march=nvptx -mcpu=sm_60 | FileCheck %s
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 | FileCheck %s
+
+; CHECK-LABEL .func test(
+define void @test(double* %dp0, double addrspace(1)* %dp1, double addrspace(3)* %dp3, double %d) {
+; CHECK: atom.add.f64
+  %r1 = call double @llvm.nvvm.atomic.load.add.f64.p0f64(double* %dp0, double %d)
+; CHECK: atom.global.add.f64
+  %r2 = call double @llvm.nvvm.atomic.load.add.f64.p1f64(double addrspace(1)* %dp1, double %d)
+; CHECK: atom.shared.add.f64
+  %ret = call double @llvm.nvvm.atomic.load.add.f64.p3f64(double addrspace(3)* %dp3, double %d)
+  ret void
+}
+
+declare double @llvm.nvvm.atomic.load.add.f64.p0f64(double* nocapture, double) #1
+declare double @llvm.nvvm.atomic.load.add.f64.p1f64(double addrspace(1)* nocapture, double) #1
+declare double @llvm.nvvm.atomic.load.add.f64.p3f64(double addrspace(3)* nocapture, double) #1
+
+attributes #1 = { argmemonly nounwind }