[NVPTX] Implement __nvvm_atom_add_gen_d builtin.

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.

Reviewers: tra

Subscribers: jholewinski, sanjoy, cfe-commits, llvm-commits, hiraditya

Differential Revision: https://reviews.llvm.org/D39638

llvm-svn: 317623

4 files changed, 39 insertions, 1 deletions

diff --git a/llvm/include/llvm/IR/IntrinsicsNVVM.td b/llvm/include/llvm/IR/IntrinsicsNVVM.td
index 7ba1a3eb2e5..249419d15d3 100644
--- a/llvm/include/llvm/IR/IntrinsicsNVVM.td
+++ b/llvm/include/llvm/IR/IntrinsicsNVVM.td
@@ -683,10 +683,15 @@ let TargetPrefix = "nvvm" in {
       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>]>;
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index 7b9acb20b75..ac4f2544fc3 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -3449,6 +3449,7 @@ bool NVPTXTargetLowering::getTgtMemIntrinsic(
   }
 
   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:
 
diff --git a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
index f745b6f6635..478f3e9d057 100644
--- a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
+++ b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
@@ -1095,6 +1095,12 @@ 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>;
@@ -1121,6 +1127,13 @@ defm INT_PTX_ATOM_ADD_S_F32 : F_ATOMIC_2<Float32Regs, ".shared", ".f32", ".add",
 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),
diff --git a/llvm/test/CodeGen/NVPTX/atomics-sm60.ll b/llvm/test/CodeGen/NVPTX/atomics-sm60.ll
new file mode 100644
index 00000000000..0b5bafb780c
--- /dev/null
+++ b/llvm/test/CodeGen/NVPTX/atomics-sm60.ll
@@ -0,0 +1,19 @@
+; 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 }