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