Home | History | Annotate | Download | only in AMDGPU 
      1 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck %s
      2 ; CHECK-DAG: flat_load_dwordx4
      3 ; CHECK-DAG: flat_load_dwordx4
      4 ; CHECK-DAG: flat_load_dwordx4
      5 ; CHECK-DAG: flat_load_dwordx4
      6 ; CHECK-DAG: ds_write2_b32
      7 ; CHECK-DAG: ds_write2_b32
      8 ; CHECK-DAG: ds_write2_b32
      9 ; CHECK-DAG: ds_write2_b32
     10 ; CHECK-DAG: ds_write2_b32
     11 ; CHECK-DAG: ds_write2_b32
     12 ; CHECK-DAG: ds_write2_b32
     13 ; CHECK-DAG: ds_write2_b32
     14 
     15 define amdgpu_kernel void @vectorize_global_local(i32 addrspace(1)* nocapture readonly %arg, i32 addrspace(3)* nocapture %arg1) {
     16 bb:
     17   %tmp = load i32, i32 addrspace(1)* %arg, align 4
     18   store i32 %tmp, i32 addrspace(3)* %arg1, align 4
     19   %tmp2 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 1
     20   %tmp3 = load i32, i32 addrspace(1)* %tmp2, align 4
     21   %tmp4 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 1
     22   store i32 %tmp3, i32 addrspace(3)* %tmp4, align 4
     23   %tmp5 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 2
     24   %tmp6 = load i32, i32 addrspace(1)* %tmp5, align 4
     25   %tmp7 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 2
     26   store i32 %tmp6, i32 addrspace(3)* %tmp7, align 4
     27   %tmp8 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 3
     28   %tmp9 = load i32, i32 addrspace(1)* %tmp8, align 4
     29   %tmp10 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 3
     30   store i32 %tmp9, i32 addrspace(3)* %tmp10, align 4
     31   %tmp11 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 4
     32   %tmp12 = load i32, i32 addrspace(1)* %tmp11, align 4
     33   %tmp13 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 4
     34   store i32 %tmp12, i32 addrspace(3)* %tmp13, align 4
     35   %tmp14 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 5
     36   %tmp15 = load i32, i32 addrspace(1)* %tmp14, align 4
     37   %tmp16 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 5
     38   store i32 %tmp15, i32 addrspace(3)* %tmp16, align 4
     39   %tmp17 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 6
     40   %tmp18 = load i32, i32 addrspace(1)* %tmp17, align 4
     41   %tmp19 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 6
     42   store i32 %tmp18, i32 addrspace(3)* %tmp19, align 4
     43   %tmp20 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 7
     44   %tmp21 = load i32, i32 addrspace(1)* %tmp20, align 4
     45   %tmp22 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 7
     46   store i32 %tmp21, i32 addrspace(3)* %tmp22, align 4
     47   %tmp23 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 8
     48   %tmp24 = load i32, i32 addrspace(1)* %tmp23, align 4
     49   %tmp25 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 8
     50   store i32 %tmp24, i32 addrspace(3)* %tmp25, align 4
     51   %tmp26 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 9
     52   %tmp27 = load i32, i32 addrspace(1)* %tmp26, align 4
     53   %tmp28 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 9
     54   store i32 %tmp27, i32 addrspace(3)* %tmp28, align 4
     55   %tmp29 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 10
     56   %tmp30 = load i32, i32 addrspace(1)* %tmp29, align 4
     57   %tmp31 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 10
     58   store i32 %tmp30, i32 addrspace(3)* %tmp31, align 4
     59   %tmp32 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 11
     60   %tmp33 = load i32, i32 addrspace(1)* %tmp32, align 4
     61   %tmp34 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 11
     62   store i32 %tmp33, i32 addrspace(3)* %tmp34, align 4
     63   %tmp35 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 12
     64   %tmp36 = load i32, i32 addrspace(1)* %tmp35, align 4
     65   %tmp37 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 12
     66   store i32 %tmp36, i32 addrspace(3)* %tmp37, align 4
     67   %tmp38 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 13
     68   %tmp39 = load i32, i32 addrspace(1)* %tmp38, align 4
     69   %tmp40 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 13
     70   store i32 %tmp39, i32 addrspace(3)* %tmp40, align 4
     71   %tmp41 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 14
     72   %tmp42 = load i32, i32 addrspace(1)* %tmp41, align 4
     73   %tmp43 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 14
     74   store i32 %tmp42, i32 addrspace(3)* %tmp43, align 4
     75   %tmp44 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 15
     76   %tmp45 = load i32, i32 addrspace(1)* %tmp44, align 4
     77   %tmp46 = getelementptr inbounds i32, i32 addrspace(3)* %arg1, i32 15
     78   store i32 %tmp45, i32 addrspace(3)* %tmp46, align 4
     79   ret void
     80 }
     81