129 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=bonaire < %s | FileCheck -check-prefixes=GCN,FUNC %s2; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GCN,FUNC %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,FUNC %s4; RUN: llc -mtriple=amdgcn -mcpu=gfx9-generic --amdhsa-code-object-version=6 < %s | FileCheck -check-prefixes=GCN,FUNC %s5 6; FUNC-LABEL: {{^}}atomic_add_ret_gds:7; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s8; GCN-DAG: s_movk_i32 m0, 0x10009; GCN: ds_add_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds10define amdgpu_kernel void @atomic_add_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {11 %val = atomicrmw volatile add ptr addrspace(2) %gds, i32 5 acq_rel12 store i32 %val, ptr addrspace(1) %out13 ret void14}15 16; FUNC-LABEL: {{^}}atomic_add_ret_gds_const_offset:17; GCN: s_movk_i32 m0, 0x8018; GCN: ds_add_rtn_u32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:20 gds19define amdgpu_kernel void @atomic_add_ret_gds_const_offset(ptr addrspace(1) %out, ptr addrspace(2) %gds) #0 {20 %gep = getelementptr i32, ptr addrspace(2) %gds, i32 521 %val = atomicrmw volatile add ptr addrspace(2) %gep, i32 5 acq_rel22 store i32 %val, ptr addrspace(1) %out23 ret void24}25 26; FUNC-LABEL: {{^}}atomic_sub_ret_gds:27; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s28; GCN-DAG: s_movk_i32 m0, 0x100029; GCN: ds_sub_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds30define amdgpu_kernel void @atomic_sub_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {31 %val = atomicrmw sub ptr addrspace(2) %gds, i32 5 acq_rel32 store i32 %val, ptr addrspace(1) %out33 ret void34}35 36; FUNC-LABEL: {{^}}atomic_and_ret_gds:37; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s38; GCN-DAG: s_movk_i32 m0, 0x100039; GCN: ds_and_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds40define amdgpu_kernel void @atomic_and_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {41 %val = atomicrmw and ptr addrspace(2) %gds, i32 5 acq_rel42 store i32 %val, ptr addrspace(1) %out43 ret void44}45 46; FUNC-LABEL: {{^}}atomic_or_ret_gds:47; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s48; GCN-DAG: s_movk_i32 m0, 0x100049; GCN: ds_or_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds50define amdgpu_kernel void @atomic_or_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {51 %val = atomicrmw or ptr addrspace(2) %gds, i32 5 acq_rel52 store i32 %val, ptr addrspace(1) %out53 ret void54}55 56; FUNC-LABEL: {{^}}atomic_xor_ret_gds:57; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s58; GCN-DAG: s_movk_i32 m0, 0x100059; GCN: ds_xor_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds60define amdgpu_kernel void @atomic_xor_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {61 %val = atomicrmw xor ptr addrspace(2) %gds, i32 5 acq_rel62 store i32 %val, ptr addrspace(1) %out63 ret void64}65 66; FUNC-LABEL: {{^}}atomic_umin_ret_gds:67; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s68; GCN-DAG: s_movk_i32 m0, 0x100069; GCN: ds_min_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds70define amdgpu_kernel void @atomic_umin_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {71 %val = atomicrmw umin ptr addrspace(2) %gds, i32 5 acq_rel72 store i32 %val, ptr addrspace(1) %out73 ret void74}75 76; FUNC-LABEL: {{^}}atomic_umax_ret_gds:77; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s78; GCN-DAG: s_movk_i32 m0, 0x100079; GCN: ds_max_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds80define amdgpu_kernel void @atomic_umax_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {81 %val = atomicrmw umax ptr addrspace(2) %gds, i32 5 acq_rel82 store i32 %val, ptr addrspace(1) %out83 ret void84}85 86; FUNC-LABEL: {{^}}atomic_imin_ret_gds:87; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s88; GCN-DAG: s_movk_i32 m0, 0x100089; GCN: ds_min_rtn_i32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds90define amdgpu_kernel void @atomic_imin_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {91 %val = atomicrmw min ptr addrspace(2) %gds, i32 5 acq_rel92 store i32 %val, ptr addrspace(1) %out93 ret void94}95 96; FUNC-LABEL: {{^}}atomic_imax_ret_gds:97; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s98; GCN-DAG: s_movk_i32 m0, 0x100099; GCN: ds_max_rtn_i32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds100define amdgpu_kernel void @atomic_imax_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {101 %val = atomicrmw max ptr addrspace(2) %gds, i32 5 acq_rel102 store i32 %val, ptr addrspace(1) %out103 ret void104}105 106; FUNC-LABEL: {{^}}atomic_xchg_ret_gds:107; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s108; GCN-DAG: s_movk_i32 m0, 0x1000109; GCN: ds_wrxchg_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds110define amdgpu_kernel void @atomic_xchg_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {111 %val = atomicrmw xchg ptr addrspace(2) %gds, i32 5 acq_rel112 store i32 %val, ptr addrspace(1) %out113 ret void114}115 116; FUNC-LABEL: {{^}}atomic_cmpxchg_ret_gds:117; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s118; GCN-DAG: s_movk_i32 m0, 0x1000119; GCN: ds_cmpst_rtn_b32 v{{[0-9]+}}, v[[OFF:[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}} gds120define amdgpu_kernel void @atomic_cmpxchg_ret_gds(ptr addrspace(1) %out, ptr addrspace(2) %gds) #1 {121 %val = cmpxchg ptr addrspace(2) %gds, i32 0, i32 1 acquire acquire122 %x = extractvalue { i32, i1 } %val, 0123 store i32 %x, ptr addrspace(1) %out124 ret void125}126 127attributes #0 = { nounwind "amdgpu-gds-size"="128" }128attributes #1 = { nounwind "amdgpu-gds-size"="4096" }129