3 files changed, 21 insertions, 12 deletions
diff --git a/llvm/test/CodeGen/NVPTX/f16-instructions.ll b/llvm/test/CodeGen/NVPTX/f16-instructions.ll
index 7788adc8698..9aa81dac126 100644
--- a/llvm/test/CodeGen/NVPTX/f16-instructions.ll
+++ b/llvm/test/CodeGen/NVPTX/f16-instructions.ll
@@ -1107,9 +1107,11 @@ define half @test_nearbyint(half %a) #0 {
 }
 
 ; CHECK-LABEL: test_round(
-; CHECK:      ld.param.b16    [[A:%h[0-9]+]], [test_round_param_0];
-; CHECK:      cvt.rni.f16.f16 [[R:%h[0-9]+]], [[A]];
-; CHECK:      st.param.b16    [func_retval0+0], [[R]];
+; CHECK:      ld.param.b16    {{.*}}, [test_round_param_0];
+; check the use of sign mask and 0.5 to implement round
+; CHECK:      and.b32 [[R:%r[0-9]+]], {{.*}}, -2147483648;
+; CHECK:      or.b32 {{.*}}, [[R]], 1056964608;
+; CHECK:      st.param.b16    [func_retval0+0], {{.*}};
 ; CHECK:      ret;
 define half @test_round(half %a) #0 {
   %r = call half @llvm.round.f16(half %a)
diff --git a/llvm/test/CodeGen/NVPTX/f16x2-instructions.ll b/llvm/test/CodeGen/NVPTX/f16x2-instructions.ll
index a8996815af4..44dda09a902 100644
--- a/llvm/test/CodeGen/NVPTX/f16x2-instructions.ll
+++ b/llvm/test/CodeGen/NVPTX/f16x2-instructions.ll
@@ -1378,12 +1378,13 @@ define <2 x half> @test_nearbyint(<2 x half> %a) #0 {
 }
 
 ; CHECK-LABEL: test_round(
-; CHECK:      ld.param.b32    [[A:%hh[0-9]+]], [test_round_param_0];
-; CHECK-DAG:  mov.b32         {[[A0:%h[0-9]+]], [[A1:%h[0-9]+]]}, [[A]];
-; CHECK-DAG:  cvt.rni.f16.f16 [[R1:%h[0-9]+]], [[A1]];
-; CHECK-DAG:  cvt.rni.f16.f16 [[R0:%h[0-9]+]], [[A0]];
-; CHECK:      mov.b32         [[R:%hh[0-9]+]], {[[R0]], [[R1]]}
-; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK:      ld.param.b32    {{.*}}, [test_round_param_0];
+; check the use of sign mask and 0.5 to implement round
+; CHECK:      and.b32 [[R1:%r[0-9]+]], {{.*}}, -2147483648;
+; CHECK:      or.b32 {{.*}}, [[R1]], 1056964608;
+; CHECK:      and.b32 [[R2:%r[0-9]+]], {{.*}}, -2147483648;
+; CHECK:      or.b32 {{.*}}, [[R2]], 1056964608;
+; CHECK:      st.param.b32    [func_retval0+0], {{.*}};
 ; CHECK:      ret;
 define <2 x half> @test_round(<2 x half> %a) #0 {
   %r = call <2 x half> @llvm.round.f16(<2 x half> %a)
diff --git a/llvm/test/CodeGen/NVPTX/math-intrins.ll b/llvm/test/CodeGen/NVPTX/math-intrins.ll
index 828a8807dcf..412b25c7a3b 100644
--- a/llvm/test/CodeGen/NVPTX/math-intrins.ll
+++ b/llvm/test/CodeGen/NVPTX/math-intrins.ll
@@ -74,21 +74,27 @@ define double @floor_double(double %a) {
 
 ; CHECK-LABEL: round_float
 define float @round_float(float %a) {
-  ; CHECK: cvt.rni.f32.f32
+; check the use of sign mask and 0.5 to implement round
+; CHECK: and.b32 [[R1:%r[0-9]+]], {{.*}}, -2147483648;
+; CHECK: or.b32 {{.*}}, [[R1]], 1056964608;
   %b = call float @llvm.round.f32(float %a)
   ret float %b
 }
 
 ; CHECK-LABEL: round_float_ftz
 define float @round_float_ftz(float %a) #1 {
-  ; CHECK: cvt.rni.ftz.f32.f32
+; check the use of sign mask and 0.5 to implement round
+; CHECK: and.b32 [[R1:%r[0-9]+]], {{.*}}, -2147483648;
+; CHECK: or.b32 {{.*}}, [[R1]], 1056964608;
   %b = call float @llvm.round.f32(float %a)
   ret float %b
 }
 
 ; CHECK-LABEL: round_double
 define double @round_double(double %a) {
-  ; CHECK: cvt.rni.f64.f64
+; check the use of 0.5 to implement round
+; CHECK: setp.lt.f64 {{.*}}, [[R:%fd[0-9]+]], 0d3FE0000000000000;
+; CHECK: add.rn.f64 {{.*}}, [[R]], 0d3FE0000000000000;
   %b = call double @llvm.round.f64(double %a)
   ret double %b
 }