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: @constexpr_gep_addrspacecast(
9; CHECK: %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
10; CHECK-NEXT: %asc = addrspacecast double addrspace(4)* %gep0 to double addrspace(3)*
11; CHECK-NEXT: store double 1.000000e+00, double addrspace(3)* %asc
12define void @constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {
13  %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
14  %asc = addrspacecast double addrspace(4)* %gep0 to double addrspace(3)*
15  store double 1.0, double addrspace(3)* %asc, align 8
16  ret void
17}
18
19; CHECK-LABEL: @constexpr_gep_gep_addrspacecast(
20; 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
21; CHECK-NEXT: %1 = addrspacecast double addrspace(3)* %gep0 to double addrspace(4)*
22; CHECK-NEXT: %gep1 = getelementptr inbounds double, double addrspace(4)* %1, i64 %idx1
23; CHECK-NEXT: %asc = addrspacecast double addrspace(4)* %gep1 to double addrspace(3)*
24; CHECK-NEXT: store double 1.000000e+00, double addrspace(3)* %asc, align 8
25define void @constexpr_gep_gep_addrspacecast(i64 %idx0, i64 %idx1) {
26  %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
27  %gep1 = getelementptr inbounds double, double addrspace(4)* %gep0, i64 %idx1
28  %asc = addrspacecast double addrspace(4)* %gep1 to double addrspace(3)*
29  store double 1.0, double addrspace(3)* %asc, align 8
30  ret void
31}
32
33; Don't crash
34; CHECK-LABEL: @vector_gep(
35; CHECK: %cast = addrspacecast <4 x [1024 x i32] addrspace(3)*> %array to <4 x [1024 x i32] addrspace(4)*>
36define amdgpu_kernel void @vector_gep(<4 x [1024 x i32] addrspace(3)*> %array) nounwind {
37  %cast = addrspacecast <4 x [1024 x i32] addrspace(3)*> %array to <4 x [1024 x i32] addrspace(4)*>
38  %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>
39  %p0 = extractelement <4 x i32 addrspace(4)*> %p, i32 0
40  %p1 = extractelement <4 x i32 addrspace(4)*> %p, i32 1
41  %p2 = extractelement <4 x i32 addrspace(4)*> %p, i32 2
42  %p3 = extractelement <4 x i32 addrspace(4)*> %p, i32 3
43  store i32 99, i32 addrspace(4)* %p0
44  store i32 99, i32 addrspace(4)* %p1
45  store i32 99, i32 addrspace(4)* %p2
46  store i32 99, i32 addrspace(4)* %p3
47  ret void
48}
49