1; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -infer-address-spaces %s | FileCheck %s
2
3; Test that pure GetElementPtr instructions not directly connected to
4; a memory operation are inferred.
5
6@lds = internal unnamed_addr addrspace(3) global [648 x double] undef, align 8
7
8; CHECK-LABEL: @simplified_constexpr_gep_addrspacecast(
9; CHECK: %gep0 = getelementptr inbounds double, double addrspace(3)* getelementptr inbounds ([648 x double], [648 x double] addrspace(3)* @lds, i64 0, i64 384), i64 %idx0
10; CHECK-NEXT: store double 1.000000e+00, double addrspace(3)* %gep0, align 8
11define void @simplified_constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {
12  %gep0 = getelementptr inbounds double, double addrspace(4)* addrspacecast (double addrspace(3)* getelementptr inbounds ([648 x double], [648 x double] addrspace(3)* @lds, i64 0, i64 384) to double addrspace(4)*), i64 %idx0
13  %asc = addrspacecast double addrspace(4)* %gep0 to double addrspace(3)*
14  store double 1.000000e+00, double addrspace(3)* %asc, align 8
15  ret void
16}
17
18; FIXME: Should be able to eliminate inner constantexpr addrspacecast.
19; CHECK-LABEL: @constexpr_gep_addrspacecast(
20; CHECK: %gep0 = getelementptr inbounds double, double addrspace(3)* addrspacecast (double addrspace(4)* getelementptr ([648 x double], [648 x double] addrspace(4)* addrspacecast ([648 x double] addrspace(3)* @lds to [648 x double] addrspace(4)*), i64 0, i64 384) to double addrspace(3)*), i64 %idx0
21; CHECK-NEXT: store double 1.000000e+00, double addrspace(3)* %gep0, align 8
22define void @constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {
23  %gep0 = getelementptr inbounds double, double addrspace(4)* getelementptr ([648 x double], [648 x double] addrspace(4)* addrspacecast ([648 x double] addrspace(3)* @lds to [648 x double] addrspace(4)*), i64 0, i64 384), i64 %idx0
24  %asc = addrspacecast double addrspace(4)* %gep0 to double addrspace(3)*
25  store double 1.0, double addrspace(3)* %asc, align 8
26  ret void
27}
28
29; CHECK-LABEL: @constexpr_gep_gep_addrspacecast(
30; CHECK: %gep0 = getelementptr inbounds double, double addrspace(3)* getelementptr inbounds ([648 x double], [648 x double] addrspace(3)* @lds, i64 0, i64 384), i64 %idx0
31; CHECK-NEXT: %gep1 = getelementptr inbounds double, double addrspace(3)* %gep0, i64 %idx1
32; CHECK-NEXT: store double 1.000000e+00, double addrspace(3)* %gep1, align 8
33define void @constexpr_gep_gep_addrspacecast(i64 %idx0, i64 %idx1) {
34  %gep0 = getelementptr inbounds double, double addrspace(4)* getelementptr ([648 x double], [648 x double] addrspace(4)* addrspacecast ([648 x double] addrspace(3)* @lds to [648 x double] addrspace(4)*), i64 0, i64 384), i64 %idx0
35  %gep1 = getelementptr inbounds double, double addrspace(4)* %gep0, i64 %idx1
36  %asc = addrspacecast double addrspace(4)* %gep1 to double addrspace(3)*
37  store double 1.0, double addrspace(3)* %asc, align 8
38  ret void
39}
40
41; Don't crash
42; CHECK-LABEL: @vector_gep(
43; CHECK: %cast = addrspacecast <4 x [1024 x i32] addrspace(3)*> %array to <4 x [1024 x i32] addrspace(4)*>
44define amdgpu_kernel void @vector_gep(<4 x [1024 x i32] addrspace(3)*> %array) nounwind {
45  %cast = addrspacecast <4 x [1024 x i32] addrspace(3)*> %array to <4 x [1024 x i32] addrspace(4)*>
46  %p = getelementptr [1024 x i32], <4 x [1024 x i32] addrspace(4)*> %cast, <4 x i16> zeroinitializer, <4 x i16> <i16 16, i16 16, i16 16, i16 16>
47  %p0 = extractelement <4 x i32 addrspace(4)*> %p, i32 0
48  %p1 = extractelement <4 x i32 addrspace(4)*> %p, i32 1
49  %p2 = extractelement <4 x i32 addrspace(4)*> %p, i32 2
50  %p3 = extractelement <4 x i32 addrspace(4)*> %p, i32 3
51  store i32 99, i32 addrspace(4)* %p0
52  store i32 99, i32 addrspace(4)* %p1
53  store i32 99, i32 addrspace(4)* %p2
54  store i32 99, i32 addrspace(4)* %p3
55  ret void
56}
57