74 lines · plain
1; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn < %s | FileCheck -check-prefixes=GCN,SI %s2; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GCN,VI %s3 4; GCN-LABEL: {{^}}no_reorder_v2f64_global_load_store:5; GCN: buffer_load_dwordx46; GCN: buffer_load_dwordx47; GCN: buffer_store_dwordx48; GCN: buffer_store_dwordx49; GCN: s_endpgm10define amdgpu_kernel void @no_reorder_v2f64_global_load_store(ptr addrspace(1) nocapture %x, ptr addrspace(1) nocapture %y) nounwind {11 %tmp1 = load <2 x double>, ptr addrspace(1) %x, align 1612 %tmp4 = load <2 x double>, ptr addrspace(1) %y, align 1613 store <2 x double> %tmp4, ptr addrspace(1) %x, align 1614 store <2 x double> %tmp1, ptr addrspace(1) %y, align 1615 ret void16}17 18; GCN-LABEL: {{^}}no_reorder_scalarized_v2f64_local_load_store:19; SI: ds_read2_b6420; SI: ds_write2_b6421 22; VI: ds_read_b12823; VI: ds_write_b12824 25; GCN: s_endpgm26define amdgpu_kernel void @no_reorder_scalarized_v2f64_local_load_store(ptr addrspace(3) nocapture %x, ptr addrspace(3) nocapture %y) nounwind {27 %tmp1 = load <2 x double>, ptr addrspace(3) %x, align 1628 %tmp4 = load <2 x double>, ptr addrspace(3) %y, align 1629 store <2 x double> %tmp4, ptr addrspace(3) %x, align 1630 store <2 x double> %tmp1, ptr addrspace(3) %y, align 1631 ret void32}33 34; GCN-LABEL: {{^}}no_reorder_split_v8i32_global_load_store:35; GCN: buffer_load_dwordx436; GCN: buffer_load_dwordx437; GCN: buffer_load_dwordx438; GCN: buffer_load_dwordx439 40 41; GCN: buffer_store_dwordx442; GCN: buffer_store_dwordx443; GCN: buffer_store_dwordx444; GCN: buffer_store_dwordx445; GCN: s_endpgm46define amdgpu_kernel void @no_reorder_split_v8i32_global_load_store(ptr addrspace(1) nocapture %x, ptr addrspace(1) nocapture %y) nounwind {47 %tmp1 = load <8 x i32>, ptr addrspace(1) %x, align 3248 %tmp4 = load <8 x i32>, ptr addrspace(1) %y, align 3249 store <8 x i32> %tmp4, ptr addrspace(1) %x, align 3250 store <8 x i32> %tmp1, ptr addrspace(1) %y, align 3251 ret void52}53 54; GCN-LABEL: {{^}}no_reorder_extload_64:55; GCN: ds_read_b6456; GCN: ds_read_b6457; GCN: ds_write_b6458; GCN-NOT: ds_read59; GCN: ds_write_b6460; GCN: s_endpgm61define amdgpu_kernel void @no_reorder_extload_64(ptr addrspace(3) nocapture %x, ptr addrspace(3) nocapture %y) nounwind {62 %tmp1 = load <2 x i32>, ptr addrspace(3) %x, align 863 %tmp4 = load <2 x i32>, ptr addrspace(3) %y, align 864 %tmp1ext = zext <2 x i32> %tmp1 to <2 x i64>65 %tmp4ext = zext <2 x i32> %tmp4 to <2 x i64>66 %tmp7 = add <2 x i64> %tmp1ext, <i64 1, i64 1>67 %tmp9 = add <2 x i64> %tmp4ext, <i64 1, i64 1>68 %trunctmp9 = trunc <2 x i64> %tmp9 to <2 x i32>69 %trunctmp7 = trunc <2 x i64> %tmp7 to <2 x i32>70 store <2 x i32> %trunctmp9, ptr addrspace(3) %x, align 871 store <2 x i32> %trunctmp7, ptr addrspace(3) %y, align 872 ret void73}74