1 ; RUN: llc -verify-machineinstrs -march=r600 -mcpu=SI -mattr=-promote-alloca < %s | FileCheck -check-prefix=SI-ALLOCA -check-prefix=SI %s
2 ; RUN: llc -verify-machineinstrs -march=r600 -mcpu=SI -mattr=+promote-alloca < %s | FileCheck -check-prefix=SI-PROMOTE -check-prefix=SI %s
4 declare i32 @llvm.SI.tid() nounwind readnone
5 declare void @llvm.AMDGPU.barrier.local() nounwind noduplicate
7 ; The required pointer calculations for the alloca'd actually requires
8 ; an add and won't be folded into the addressing, which fails with a
9 ; 64-bit pointer add. This should work since private pointers should
12 ; SI-LABEL: @test_private_array_ptr_calc:
14 ; SI: V_ADD_I32_e32 [[PTRREG:v[0-9]+]]
16 ; SI-ALLOCA: V_MOVRELD_B32_e32 {{v[0-9]+}}, [[PTRREG]]
18 ; FIXME: The AMDGPUPromoteAlloca pass should be able to convert this
19 ; alloca to a vector. It currently fails because it does not know how
21 ; getelementptr [4 x i32]* %alloca, i32 1, i32 %b
23 ; SI-PROMOTE: DS_WRITE_B32 {{v[0-9]+}}, [[PTRREG]]
24 define void @test_private_array_ptr_calc(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %inA, i32 addrspace(1)* noalias %inB) {
25 %alloca = alloca [4 x i32], i32 4, align 16
26 %tid = call i32 @llvm.SI.tid() readnone
27 %a_ptr = getelementptr i32 addrspace(1)* %inA, i32 %tid
28 %b_ptr = getelementptr i32 addrspace(1)* %inB, i32 %tid
29 %a = load i32 addrspace(1)* %a_ptr
30 %b = load i32 addrspace(1)* %b_ptr
31 %result = add i32 %a, %b
32 %alloca_ptr = getelementptr [4 x i32]* %alloca, i32 1, i32 %b
33 store i32 %result, i32* %alloca_ptr, align 4
35 call void @llvm.AMDGPU.barrier.local() nounwind noduplicate
36 %reload = load i32* %alloca_ptr, align 4
37 %out_ptr = getelementptr i32 addrspace(1)* %out, i32 %tid
38 store i32 %reload, i32 addrspace(1)* %out_ptr, align 4