[NVPTX] Add __nvvm_* intrinsics as Clang builtins

Fixes bug 13354.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@167647 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/include/clang/Basic/BuiltinsNVPTX.def b/include/clang/Basic/BuiltinsNVPTX.def
index f90a43f..2fde59b 100644
--- a/include/clang/Basic/BuiltinsNVPTX.def
+++ b/include/clang/Basic/BuiltinsNVPTX.def
@@ -14,6 +14,7 @@
 
 // The format of this database matches clang/Basic/Builtins.def.
 
+// Builtins retained from previous PTX back-end
 BUILTIN(__builtin_ptx_read_tid_x, "i", "nc")
 BUILTIN(__builtin_ptx_read_tid_y, "i", "nc")
 BUILTIN(__builtin_ptx_read_tid_z, "i", "nc")
@@ -59,4 +60,249 @@
 BUILTIN(__builtin_ptx_bar_sync, "vi", "n")
 
 
+// Builtins exposed as part of NVVM
+BUILTIN(__syncthreads, "v", "nc")
+BUILTIN(__nvvm_bar0, "v", "nc")
+BUILTIN(__nvvm_bar0_popc, "ii", "nc")
+BUILTIN(__nvvm_bar0_and, "ii", "nc")
+BUILTIN(__nvvm_bar0_or, "ii", "nc")
+BUILTIN(__nvvm_membar_cta, "v", "n")
+BUILTIN(__nvvm_membar_gl, "v", "n")
+BUILTIN(__nvvm_membar_sys, "v", "n")
+BUILTIN(__nvvm_popc_i, "ii", "n")
+BUILTIN(__nvvm_popc_ll, "iLLi", "n")
+BUILTIN(__nvvm_prmt, "UiUiUiUi", "nc")
+BUILTIN(__nvvm_min_i, "iii", "nc")
+BUILTIN(__nvvm_min_ui, "UiUiUi", "nc")
+BUILTIN(__nvvm_min_ll, "LLiLLiLLi", "nc")
+BUILTIN(__nvvm_min_ull, "ULLiULLiULLi", "nc")
+BUILTIN(__nvvm_max_i, "iii", "nc")
+BUILTIN(__nvvm_max_ui, "UiUiUi", "nc")
+BUILTIN(__nvvm_max_ll, "LLiLLiLLi", "nc")
+BUILTIN(__nvvm_max_ull, "ULLiULLiULLi", "nc")
+BUILTIN(__nvvm_mulhi_i, "iii", "nc")
+BUILTIN(__nvvm_mulhi_ui, "UiUiUi", "nc")
+BUILTIN(__nvvm_mulhi_ll, "LLiLLiLLi", "nc")
+BUILTIN(__nvvm_mulhi_ull, "ULLiULLiULLi", "nc")
+BUILTIN(__nvvm_mul24_i, "iii", "nc")
+BUILTIN(__nvvm_mul24_ui, "UiUiUi", "nc")
+BUILTIN(__nvvm_brev32, "UiUi", "nc")
+BUILTIN(__nvvm_brev64, "ULLiULLi", "nc")
+BUILTIN(__nvvm_sad_i, "iiii", "nc")
+BUILTIN(__nvvm_sad_ui, "UiUiUiUi", "nc")
+BUILTIN(__nvvm_abs_i, "ii", "nc")
+BUILTIN(__nvvm_abs_ll, "LiLi", "nc")
+BUILTIN(__nvvm_floor_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_floor_f, "ff", "nc")
+BUILTIN(__nvvm_floor_d, "dd", "nc")
+BUILTIN(__nvvm_fabs_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_fabs_f, "ff", "nc")
+BUILTIN(__nvvm_fabs_d, "dd", "nc")
+BUILTIN(__nvvm_rcp_approx_ftz_d, "dd", "nc")
+BUILTIN(__nvvm_fmin_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_fmin_f, "fff", "nc")
+BUILTIN(__nvvm_fmax_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_fmax_f, "fff", "nc")
+BUILTIN(__nvvm_rsqrt_approx_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_rsqrt_approx_f, "ff", "nc")
+BUILTIN(__nvvm_fmin_d, "ddd", "nc")
+BUILTIN(__nvvm_fmax_d, "ddd", "nc")
+BUILTIN(__nvvm_rsqrt_approx_d, "dd", "nc")
+BUILTIN(__nvvm_ceil_d, "dd", "nc")
+BUILTIN(__nvvm_trunc_d, "dd", "nc")
+BUILTIN(__nvvm_round_d, "dd", "nc")
+BUILTIN(__nvvm_ex2_approx_d, "dd", "nc")
+BUILTIN(__nvvm_lg2_approx_d, "dd", "nc")
+BUILTIN(__nvvm_round_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_round_f, "ff", "nc")
+BUILTIN(__nvvm_ex2_approx_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_ex2_approx_f, "ff", "nc")
+BUILTIN(__nvvm_lg2_approx_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_lg2_approx_f, "ff", "nc")
+BUILTIN(__nvvm_sin_approx_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_sin_approx_f, "ff", "nc")
+BUILTIN(__nvvm_cos_approx_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_cos_approx_f, "ff", "nc")
+BUILTIN(__nvvm_trunc_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_trunc_f, "ff", "nc")
+BUILTIN(__nvvm_ceil_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_ceil_f, "ff", "nc")
+BUILTIN(__nvvm_saturate_d, "dd", "nc")
+BUILTIN(__nvvm_saturate_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_saturate_f, "ff", "nc")
+BUILTIN(__nvvm_fma_rn_ftz_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rn_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rz_ftz_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rz_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rm_ftz_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rm_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rp_ftz_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rp_f, "ffff", "nc")
+BUILTIN(__nvvm_fma_rn_d, "dddd", "nc")
+BUILTIN(__nvvm_fma_rz_d, "dddd", "nc")
+BUILTIN(__nvvm_fma_rm_d, "dddd", "nc")
+BUILTIN(__nvvm_fma_rp_d, "dddd", "nc")
+BUILTIN(__nvvm_div_approx_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_div_approx_f, "fff", "nc")
+BUILTIN(__nvvm_div_rn_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_div_rn_f, "fff", "nc")
+BUILTIN(__nvvm_div_rz_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_div_rz_f, "fff", "nc")
+BUILTIN(__nvvm_div_rm_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_div_rm_f, "fff", "nc")
+BUILTIN(__nvvm_div_rp_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_div_rp_f, "fff", "nc")
+BUILTIN(__nvvm_rcp_rn_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_rcp_rn_f, "ff", "nc")
+BUILTIN(__nvvm_rcp_rz_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_rcp_rz_f, "ff", "nc")
+BUILTIN(__nvvm_rcp_rm_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_rcp_rm_f, "ff", "nc")
+BUILTIN(__nvvm_rcp_rp_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_rcp_rp_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rn_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rn_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rz_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rz_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rm_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rm_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rp_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_rp_f, "ff", "nc")
+BUILTIN(__nvvm_div_rn_d, "ddd", "nc")
+BUILTIN(__nvvm_div_rz_d, "ddd", "nc")
+BUILTIN(__nvvm_div_rm_d, "ddd", "nc")
+BUILTIN(__nvvm_div_rp_d, "ddd", "nc")
+BUILTIN(__nvvm_rcp_rn_d, "dd", "nc")
+BUILTIN(__nvvm_rcp_rz_d, "dd", "nc")
+BUILTIN(__nvvm_rcp_rm_d, "dd", "nc")
+BUILTIN(__nvvm_rcp_rp_d, "dd", "nc")
+BUILTIN(__nvvm_sqrt_rn_d, "dd", "nc")
+BUILTIN(__nvvm_sqrt_rz_d, "dd", "nc")
+BUILTIN(__nvvm_sqrt_rm_d, "dd", "nc")
+BUILTIN(__nvvm_sqrt_rp_d, "dd", "nc")
+BUILTIN(__nvvm_sqrt_approx_ftz_f, "ff", "nc")
+BUILTIN(__nvvm_sqrt_approx_f, "ff", "nc")
+BUILTIN(__nvvm_add_rn_d, "ddd", "nc")
+BUILTIN(__nvvm_add_rz_d, "ddd", "nc")
+BUILTIN(__nvvm_add_rm_d, "ddd", "nc")
+BUILTIN(__nvvm_add_rp_d, "ddd", "nc")
+BUILTIN(__nvvm_mul_rn_d, "ddd", "nc")
+BUILTIN(__nvvm_mul_rz_d, "ddd", "nc")
+BUILTIN(__nvvm_mul_rm_d, "ddd", "nc")
+BUILTIN(__nvvm_mul_rp_d, "ddd", "nc")
+BUILTIN(__nvvm_add_rm_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_add_rm_f, "fff", "nc")
+BUILTIN(__nvvm_add_rp_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_add_rp_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rm_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rm_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rp_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rp_f, "fff", "nc")
+BUILTIN(__nvvm_add_rn_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_add_rn_f, "fff", "nc")
+BUILTIN(__nvvm_add_rz_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_add_rz_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rn_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rn_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rz_ftz_f, "fff", "nc")
+BUILTIN(__nvvm_mul_rz_f, "fff", "nc")
+BUILTIN(__nvvm_d2f_rn_ftz, "fd", "nc")
+BUILTIN(__nvvm_d2f_rn, "fd", "nc")
+BUILTIN(__nvvm_d2f_rz_ftz, "fd", "nc")
+BUILTIN(__nvvm_d2f_rz, "fd", "nc")
+BUILTIN(__nvvm_d2f_rm_ftz, "fd", "nc")
+BUILTIN(__nvvm_d2f_rm, "fd", "nc")
+BUILTIN(__nvvm_d2f_rp_ftz, "fd", "nc")
+BUILTIN(__nvvm_d2f_rp, "fd", "nc")
+BUILTIN(__nvvm_d2i_rn, "id", "nc")
+BUILTIN(__nvvm_d2i_rz, "id", "nc")
+BUILTIN(__nvvm_d2i_rm, "id", "nc")
+BUILTIN(__nvvm_d2i_rp, "id", "nc")
+BUILTIN(__nvvm_d2ui_rn, "Uid", "nc")
+BUILTIN(__nvvm_d2ui_rz, "Uid", "nc")
+BUILTIN(__nvvm_d2ui_rm, "Uid", "nc")
+BUILTIN(__nvvm_d2ui_rp, "Uid", "nc")
+BUILTIN(__nvvm_i2d_rn, "di", "nc")
+BUILTIN(__nvvm_i2d_rz, "di", "nc")
+BUILTIN(__nvvm_i2d_rm, "di", "nc")
+BUILTIN(__nvvm_i2d_rp, "di", "nc")
+BUILTIN(__nvvm_ui2d_rn, "dUi", "nc")
+BUILTIN(__nvvm_ui2d_rz, "dUi", "nc")
+BUILTIN(__nvvm_ui2d_rm, "dUi", "nc")
+BUILTIN(__nvvm_ui2d_rp, "dUi", "nc")
+BUILTIN(__nvvm_f2i_rn_ftz, "if", "nc")
+BUILTIN(__nvvm_f2i_rn, "if", "nc")
+BUILTIN(__nvvm_f2i_rz_ftz, "if", "nc")
+BUILTIN(__nvvm_f2i_rz, "if", "nc")
+BUILTIN(__nvvm_f2i_rm_ftz, "if", "nc")
+BUILTIN(__nvvm_f2i_rm, "if", "nc")
+BUILTIN(__nvvm_f2i_rp_ftz, "if", "nc")
+BUILTIN(__nvvm_f2i_rp, "if", "nc")
+BUILTIN(__nvvm_f2ui_rn_ftz, "Uif", "nc")
+BUILTIN(__nvvm_f2ui_rn, "Uif", "nc")
+BUILTIN(__nvvm_f2ui_rz_ftz, "Uif", "nc")
+BUILTIN(__nvvm_f2ui_rz, "Uif", "nc")
+BUILTIN(__nvvm_f2ui_rm_ftz, "Uif", "nc")
+BUILTIN(__nvvm_f2ui_rm, "Uif", "nc")
+BUILTIN(__nvvm_f2ui_rp_ftz, "Uif", "nc")
+BUILTIN(__nvvm_f2ui_rp, "Uif", "nc")
+BUILTIN(__nvvm_i2f_rn, "fi", "nc")
+BUILTIN(__nvvm_i2f_rz, "fi", "nc")
+BUILTIN(__nvvm_i2f_rm, "fi", "nc")
+BUILTIN(__nvvm_i2f_rp, "fi", "nc")
+BUILTIN(__nvvm_ui2f_rn, "fUi", "nc")
+BUILTIN(__nvvm_ui2f_rz, "fUi", "nc")
+BUILTIN(__nvvm_ui2f_rm, "fUi", "nc")
+BUILTIN(__nvvm_ui2f_rp, "fUi", "nc")
+BUILTIN(__nvvm_lohi_i2d, "dii", "nc")
+BUILTIN(__nvvm_d2i_lo, "id", "nc")
+BUILTIN(__nvvm_d2i_hi, "id", "nc")
+BUILTIN(__nvvm_f2ll_rn_ftz, "LLif", "nc")
+BUILTIN(__nvvm_f2ll_rn, "LLif", "nc")
+BUILTIN(__nvvm_f2ll_rz_ftz, "LLif", "nc")
+BUILTIN(__nvvm_f2ll_rz, "LLif", "nc")
+BUILTIN(__nvvm_f2ll_rm_ftz, "LLif", "nc")
+BUILTIN(__nvvm_f2ll_rm, "LLif", "nc")
+BUILTIN(__nvvm_f2ll_rp_ftz, "LLif", "nc")
+BUILTIN(__nvvm_f2ll_rp, "LLif", "nc")
+BUILTIN(__nvvm_f2ull_rn_ftz, "ULLif", "nc")
+BUILTIN(__nvvm_f2ull_rn, "ULLif", "nc")
+BUILTIN(__nvvm_f2ull_rz_ftz, "ULLif", "nc")
+BUILTIN(__nvvm_f2ull_rz, "ULLif", "nc")
+BUILTIN(__nvvm_f2ull_rm_ftz, "ULLif", "nc")
+BUILTIN(__nvvm_f2ull_rm, "ULLif", "nc")
+BUILTIN(__nvvm_f2ull_rp_ftz, "ULLif", "nc")
+BUILTIN(__nvvm_f2ull_rp, "ULLif", "nc")
+BUILTIN(__nvvm_d2ll_rn, "LLid", "nc")
+BUILTIN(__nvvm_d2ll_rz, "LLid", "nc")
+BUILTIN(__nvvm_d2ll_rm, "LLid", "nc")
+BUILTIN(__nvvm_d2ll_rp, "LLid", "nc")
+BUILTIN(__nvvm_d2ull_rn, "ULLid", "nc")
+BUILTIN(__nvvm_d2ull_rz, "ULLid", "nc")
+BUILTIN(__nvvm_d2ull_rm, "ULLid", "nc")
+BUILTIN(__nvvm_d2ull_rp, "ULLid", "nc")
+BUILTIN(__nvvm_ll2f_rn, "fLLi", "nc")
+BUILTIN(__nvvm_ll2f_rz, "fLLi", "nc")
+BUILTIN(__nvvm_ll2f_rm, "fLLi", "nc")
+BUILTIN(__nvvm_ll2f_rp, "fLLi", "nc")
+BUILTIN(__nvvm_ull2f_rn, "fULLi", "nc")
+BUILTIN(__nvvm_ull2f_rz, "fULLi", "nc")
+BUILTIN(__nvvm_ull2f_rm, "fULLi", "nc")
+BUILTIN(__nvvm_ull2f_rp, "fULLi", "nc")
+BUILTIN(__nvvm_ll2d_rn, "dLLi", "nc")
+BUILTIN(__nvvm_ll2d_rz, "dLLi", "nc")
+BUILTIN(__nvvm_ll2d_rm, "dLLi", "nc")
+BUILTIN(__nvvm_ll2d_rp, "dLLi", "nc")
+BUILTIN(__nvvm_ull2d_rn, "dULLi", "nc")
+BUILTIN(__nvvm_ull2d_rz, "dULLi", "nc")
+BUILTIN(__nvvm_ull2d_rm, "dULLi", "nc")
+BUILTIN(__nvvm_ull2d_rp, "dULLi", "nc")
+BUILTIN(__nvvm_f2h_rn_ftz, "Usf", "nc")
+BUILTIN(__nvvm_f2h_rn, "Usf", "nc")
+BUILTIN(__nvvm_h2f, "fUs", "nc")
+BUILTIN(__nvvm_bitcast_i2f, "fi", "nc")
+BUILTIN(__nvvm_bitcast_f2i, "if", "nc")
+BUILTIN(__nvvm_bitcast_ll2d, "dLLi", "nc")
+BUILTIN(__nvvm_bitcast_d2ll, "LLid", "nc")
+
 #undef BUILTIN
diff --git a/test/CodeGen/builtins-nvptx.c b/test/CodeGen/builtins-nvptx.c
index fa6b14c..2c7e0c1 100644
--- a/test/CodeGen/builtins-nvptx.c
+++ b/test/CodeGen/builtins-nvptx.c
@@ -1,8 +1,15 @@
-// RUN: %clang_cc1 -triple nvptx-unknown-unknown -emit-llvm -o %t %s
-// RUN: %clang_cc1 -triple nvptx64-unknown-unknown -emit-llvm -o %t %s
+// REQUIRES: nvptx-registered-target
+// REQUIRES: nvptx64-registered-target
+// RUN: %clang_cc1 -triple nvptx-unknown-unknown -S -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple nvptx64-unknown-unknown -S -emit-llvm -o - %s | FileCheck %s
 
 int read_tid() {
 
+// CHECK: call i32 @llvm.ptx.read.tid.x()
+// CHECK: call i32 @llvm.ptx.read.tid.y()
+// CHECK: call i32 @llvm.ptx.read.tid.z()
+// CHECK: call i32 @llvm.ptx.read.tid.w()
+
   int x = __builtin_ptx_read_tid_x();
   int y = __builtin_ptx_read_tid_y();
   int z = __builtin_ptx_read_tid_z();
@@ -14,6 +21,11 @@
 
 int read_ntid() {
 
+// CHECK: call i32 @llvm.ptx.read.ntid.x()
+// CHECK: call i32 @llvm.ptx.read.ntid.y()
+// CHECK: call i32 @llvm.ptx.read.ntid.z()
+// CHECK: call i32 @llvm.ptx.read.ntid.w()
+
   int x = __builtin_ptx_read_ntid_x();
   int y = __builtin_ptx_read_ntid_y();
   int z = __builtin_ptx_read_ntid_z();
@@ -25,6 +37,11 @@
 
 int read_ctaid() {
 
+// CHECK: call i32 @llvm.ptx.read.ctaid.x()
+// CHECK: call i32 @llvm.ptx.read.ctaid.y()
+// CHECK: call i32 @llvm.ptx.read.ctaid.z()
+// CHECK: call i32 @llvm.ptx.read.ctaid.w()
+
   int x = __builtin_ptx_read_ctaid_x();
   int y = __builtin_ptx_read_ctaid_y();
   int z = __builtin_ptx_read_ctaid_z();
@@ -36,6 +53,11 @@
 
 int read_nctaid() {
 
+// CHECK: call i32 @llvm.ptx.read.nctaid.x()
+// CHECK: call i32 @llvm.ptx.read.nctaid.y()
+// CHECK: call i32 @llvm.ptx.read.nctaid.z()
+// CHECK: call i32 @llvm.ptx.read.nctaid.w()
+
   int x = __builtin_ptx_read_nctaid_x();
   int y = __builtin_ptx_read_nctaid_y();
   int z = __builtin_ptx_read_nctaid_z();
@@ -47,6 +69,13 @@
 
 int read_ids() {
 
+// CHECK: call i32 @llvm.ptx.read.laneid()
+// CHECK: call i32 @llvm.ptx.read.warpid()
+// CHECK: call i32 @llvm.ptx.read.nwarpid()
+// CHECK: call i32 @llvm.ptx.read.smid()
+// CHECK: call i32 @llvm.ptx.read.nsmid()
+// CHECK: call i32 @llvm.ptx.read.gridid()
+
   int a = __builtin_ptx_read_laneid();
   int b = __builtin_ptx_read_warpid();
   int c = __builtin_ptx_read_nwarpid();
@@ -60,6 +89,12 @@
 
 int read_lanemasks() {
 
+// CHECK: call i32 @llvm.ptx.read.lanemask.eq()
+// CHECK: call i32 @llvm.ptx.read.lanemask.le()
+// CHECK: call i32 @llvm.ptx.read.lanemask.lt()
+// CHECK: call i32 @llvm.ptx.read.lanemask.ge()
+// CHECK: call i32 @llvm.ptx.read.lanemask.gt()
+
   int a = __builtin_ptx_read_lanemask_eq();
   int b = __builtin_ptx_read_lanemask_le();
   int c = __builtin_ptx_read_lanemask_lt();
@@ -73,6 +108,9 @@
 
 long read_clocks() {
 
+// CHECK: call i32 @llvm.ptx.read.clock()
+// CHECK: call i64 @llvm.ptx.read.clock64()
+
   int a = __builtin_ptx_read_clock();
   long b = __builtin_ptx_read_clock64();
 
@@ -82,6 +120,11 @@
 
 int read_pms() {
 
+// CHECK: call i32 @llvm.ptx.read.pm0()
+// CHECK: call i32 @llvm.ptx.read.pm1()
+// CHECK: call i32 @llvm.ptx.read.pm2()
+// CHECK: call i32 @llvm.ptx.read.pm3()
+
   int a = __builtin_ptx_read_pm0();
   int b = __builtin_ptx_read_pm1();
   int c = __builtin_ptx_read_pm2();
@@ -93,6 +136,33 @@
 
 void sync() {
 
+// CHECK: call void @llvm.ptx.bar.sync(i32 0)
+
   __builtin_ptx_bar_sync(0);
 
 }
+
+
+// NVVM intrinsics
+
+// The idea is not to test all intrinsics, just that Clang is recognizing the
+// builtins defined in BuiltinsNVPTX.def
+void nvvm_math(float f1, float f2, double d1, double d2) {
+// CHECK: call float @llvm.nvvm.fmax.f
+  float t1 = __nvvm_fmax_f(f1, f2);
+// CHECK: call float @llvm.nvvm.fmin.f
+  float t2 = __nvvm_fmin_f(f1, f2);
+// CHECK: call float @llvm.nvvm.sqrt.rn.f
+  float t3 = __nvvm_sqrt_rn_f(f1);
+// CHECK: call float @llvm.nvvm.rcp.rn.f
+  float t4 = __nvvm_rcp_rn_f(f2);
+
+// CHECK: call double @llvm.nvvm.fmax.d
+  double td1 = __nvvm_fmax_d(d1, d2);
+// CHECK: call double @llvm.nvvm.fmin.d
+  double td2 = __nvvm_fmin_d(d1, d2);
+// CHECK: call double @llvm.nvvm.sqrt.rn.d
+  double td3 = __nvvm_sqrt_rn_d(d1);
+// CHECK: call double @llvm.nvvm.rcp.rn.d
+  double td4 = __nvvm_rcp_rn_d(d2);
+}