1; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s
2; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 | %ptxas-verify %}
3
4target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
5target triple = "nvptx64-nvidia-cuda"
6
7; Verify that both %input and %output are converted to global pointers and then
8; addrspacecast'ed back to the original type.
9define void @kernel(float* %input, float* %output) {
10; CHECK-LABEL: .visible .entry kernel(
11; CHECK: cvta.to.global.u64
12; CHECK: cvta.to.global.u64
13  %1 = load float, float* %input, align 4
14; CHECK: ld.global.f32
15  store float %1, float* %output, align 4
16; CHECK: st.global.f32
17  ret void
18}
19
20define void @kernel2(float addrspace(1)* %input, float addrspace(1)* %output) {
21; CHECK-LABEL: .visible .entry kernel2(
22; CHECK-NOT: cvta.to.global.u64
23  %1 = load float, float addrspace(1)* %input, align 4
24; CHECK: ld.global.f32
25  store float %1, float addrspace(1)* %output, align 4
26; CHECK: st.global.f32
27  ret void
28}
29
30%struct.S = type { i32*, i32* }
31
32define void @ptr_in_byval_kernel(%struct.S* byval(%struct.S) %input, i32* %output) {
33; CHECK-LABEL: .visible .entry ptr_in_byval_kernel(
34; CHECK: ld.param.u64 	%[[optr:rd.*]], [ptr_in_byval_kernel_param_1]
35; CHECK: cvta.to.global.u64 %[[optr_g:.*]], %[[optr]];
36; CHECK: ld.param.u64 	%[[iptr:rd.*]], [ptr_in_byval_kernel_param_0+8]
37; CHECK: cvta.to.global.u64 %[[iptr_g:.*]], %[[iptr]];
38  %b_ptr = getelementptr inbounds %struct.S, %struct.S* %input, i64 0, i32 1
39  %b = load i32*, i32** %b_ptr, align 8
40  %v = load i32, i32* %b, align 4
41; CHECK: ld.global.u32 %[[val:.*]], [%[[iptr_g]]]
42  store i32 %v, i32* %output, align 4
43; CHECK: st.global.u32 [%[[optr_g]]], %[[val]]
44  ret void
45}
46
47; Regular functions lower byval arguments differently. We need to make
48; sure that we're loading byval argument data using [symbol+offset].
49; There's also no assumption that all pointers within are in global space.
50define void @ptr_in_byval_func(%struct.S* byval(%struct.S) %input, i32* %output) {
51; CHECK-LABEL: .visible .func ptr_in_byval_func(
52; CHECK: ld.param.u64 	%[[optr:rd.*]], [ptr_in_byval_func_param_1]
53; CHECK: ld.param.u64 	%[[iptr:rd.*]], [ptr_in_byval_func_param_0+8]
54  %b_ptr = getelementptr inbounds %struct.S, %struct.S* %input, i64 0, i32 1
55  %b = load i32*, i32** %b_ptr, align 8
56  %v = load i32, i32* %b, align 4
57; CHECK: ld.u32 %[[val:.*]], [%[[iptr]]]
58  store i32 %v, i32* %output, align 4
59; CHECK: st.u32 [%[[optr]]], %[[val]]
60  ret void
61}
62
63!nvvm.annotations = !{!0, !1, !2}
64!0 = !{void (float*, float*)* @kernel, !"kernel", i32 1}
65!1 = !{void (float addrspace(1)*, float addrspace(1)*)* @kernel2, !"kernel", i32 1}
66!2 = !{void (%struct.S*, i32*)* @ptr_in_byval_kernel, !"kernel", i32 1}
67