1 ; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=FUNC -check-prefix=SI %s
3 ; FIXME: This currently doesn't do a great job of clustering the
4 ; loads, which end up with extra moves between them. Right now, it
5 ; seems the only things areLoadsFromSameBasePtr is accomplishing is
6 ; ordering the loads so that the lower address loads come first.
8 ; FUNC-LABEL: {{^}}cluster_global_arg_loads:
9 ; SI-DAG: buffer_load_dword [[REG0:v[0-9]+]], off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
10 ; SI-DAG: buffer_load_dword [[REG1:v[0-9]+]], off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:8
11 ; SI: buffer_store_dword [[REG0]]
12 ; SI: buffer_store_dword [[REG1]]
13 define amdgpu_kernel void @cluster_global_arg_loads(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 addrspace(1)* %ptr) #0 {
14 %load0 = load i32, i32 addrspace(1)* %ptr, align 4
15 %gep = getelementptr i32, i32 addrspace(1)* %ptr, i32 2
16 %load1 = load i32, i32 addrspace(1)* %gep, align 4
17 store i32 %load0, i32 addrspace(1)* %out0, align 4
18 store i32 %load1, i32 addrspace(1)* %out1, align 4
22 ; Test for a crach in SIInstrInfo::areLoadsFromSameBasePtr() when checking
23 ; an MUBUF load which does not have a vaddr operand.
24 ; FUNC-LABEL: {{^}}same_base_ptr_crash:
25 ; SI: buffer_load_dword
26 ; SI: buffer_load_dword
27 define amdgpu_kernel void @same_base_ptr_crash(i32 addrspace(1)* %out, i32 addrspace(1)* %in, i32 %offset) {
29 %out1 = getelementptr i32, i32 addrspace(1)* %out, i32 %offset
30 %tmp0 = load i32, i32 addrspace(1)* %out
31 %tmp1 = load i32, i32 addrspace(1)* %out1
32 %tmp2 = add i32 %tmp0, %tmp1
33 store i32 %tmp2, i32 addrspace(1)* %out
37 attributes #0 = { nounwind }
38 attributes #1 = { nounwind readnone }