1; RUN: llc -O0 -mtriple=amdgcn--amdhsa -march=amdgcn -amdgpu-spill-sgpr-to-vgpr=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=VMEM -check-prefix=GCN %s 2; RUN: llc -O0 -mtriple=amdgcn--amdhsa -march=amdgcn -amdgpu-spill-sgpr-to-vgpr=1 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=VGPR -check-prefix=GCN %s 3 4; Verify registers used for tracking exec mask changes when all 5; registers are spilled at the end of the block. The SGPR spill 6; placement relative to the exec modifications are important. 7 8; FIXME: This checks with SGPR to VGPR spilling disabled, but this may 9; not work correctly in cases where no workitems take a branch. 10 11 12; GCN-LABEL: {{^}}divergent_if_endif: 13; VGPR: workitem_private_segment_byte_size = 12{{$}} 14 15 16; GCN: {{^}}; %bb.0: 17; GCN: s_mov_b32 m0, -1 18; GCN: ds_read_b32 [[LOAD0:v[0-9]+]] 19 20; GCN: v_cmp_eq_u32_e64 [[CMP0:s\[[0-9]+:[0-9]\]]], s{{[0-9]+}}, v0 21; GCN: s_mov_b64 s{{\[}}[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]{{\]}}, exec 22; GCN: s_and_b64 s{{\[}}[[ANDEXEC_LO:[0-9]+]]:[[ANDEXEC_HI:[0-9]+]]{{\]}}, s{{\[}}[[SAVEEXEC_LO]]:[[SAVEEXEC_HI]]{{\]}}, [[CMP0]] 23 24; Spill load 25; GCN: buffer_store_dword [[LOAD0]], off, s[0:3], s7 offset:[[LOAD0_OFFSET:[0-9]+]] ; 4-byte Folded Spill 26 27; Spill saved exec 28; VGPR: v_writelane_b32 [[SPILL_VGPR:v[0-9]+]], s[[SAVEEXEC_LO]], [[SAVEEXEC_LO_LANE:[0-9]+]] 29; VGPR: v_writelane_b32 [[SPILL_VGPR]], s[[SAVEEXEC_HI]], [[SAVEEXEC_HI_LANE:[0-9]+]] 30 31; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_LO:[0-9]+]], s[[SAVEEXEC_LO]] 32; VMEM: buffer_store_dword v[[V_SAVEEXEC_LO]], off, s[0:3], s7 offset:20 ; 4-byte Folded Spill 33; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_HI:[0-9]+]], s[[SAVEEXEC_HI]] 34; VMEM: buffer_store_dword v[[V_SAVEEXEC_HI]], off, s[0:3], s7 offset:24 ; 4-byte Folded Spill 35 36; GCN: s_mov_b64 exec, s{{\[}}[[ANDEXEC_LO]]:[[ANDEXEC_HI]]{{\]}} 37 38; GCN: mask branch [[ENDIF:BB[0-9]+_[0-9]+]] 39 40; GCN: {{^}}BB{{[0-9]+}}_1: ; %if 41; GCN: s_mov_b32 m0, -1 42; GCN: ds_read_b32 [[LOAD1:v[0-9]+]] 43; GCN: buffer_load_dword [[RELOAD_LOAD0:v[0-9]+]], off, s[0:3], s7 offset:[[LOAD0_OFFSET]] ; 4-byte Folded Reload 44; GCN: s_waitcnt vmcnt(0) lgkmcnt(0) 45 46 47; Spill val register 48; GCN: v_add_i32_e32 [[VAL:v[0-9]+]], vcc, [[LOAD1]], [[RELOAD_LOAD0]] 49; GCN: buffer_store_dword [[VAL]], off, s[0:3], s7 offset:[[VAL_OFFSET:[0-9]+]] ; 4-byte Folded Spill 50 51; VMEM: [[ENDIF]]: 52 53; Reload and restore exec mask 54; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_LO_LANE]] 55; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_HI_LANE]] 56 57 58 59; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_LO:[0-9]+]], off, s[0:3], s7 offset:20 ; 4-byte Folded Reload 60; VMEM: s_waitcnt vmcnt(0) 61; VMEM: v_readfirstlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], v[[V_RELOAD_SAVEEXEC_LO]] 62 63; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_HI:[0-9]+]], off, s[0:3], s7 offset:24 ; 4-byte Folded Reload 64; VMEM: s_waitcnt vmcnt(0) 65; VMEM: v_readfirstlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], v[[V_RELOAD_SAVEEXEC_HI]] 66 67; GCN: s_or_b64 exec, exec, s{{\[}}[[S_RELOAD_SAVEEXEC_LO]]:[[S_RELOAD_SAVEEXEC_HI]]{{\]}} 68 69; Restore val 70; GCN: buffer_load_dword [[RELOAD_VAL:v[0-9]+]], off, s[0:3], s7 offset:[[VAL_OFFSET]] ; 4-byte Folded Reload 71 72; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RELOAD_VAL]] 73define amdgpu_kernel void @divergent_if_endif(i32 addrspace(1)* %out) #0 { 74entry: 75 %tid = call i32 @llvm.amdgcn.workitem.id.x() 76 %load0 = load volatile i32, i32 addrspace(3)* undef 77 %cmp0 = icmp eq i32 %tid, 0 78 br i1 %cmp0, label %if, label %endif 79 80if: 81 %load1 = load volatile i32, i32 addrspace(3)* undef 82 %val = add i32 %load0, %load1 83 br label %endif 84 85endif: 86 %tmp4 = phi i32 [ %val, %if ], [ 0, %entry ] 87 store i32 %tmp4, i32 addrspace(1)* %out 88 ret void 89} 90 91; GCN-LABEL: {{^}}divergent_loop: 92; VGPR: workitem_private_segment_byte_size = 16{{$}} 93 94; GCN: {{^}}; %bb.0: 95 96; GCN: s_mov_b32 m0, -1 97; GCN: ds_read_b32 [[LOAD0:v[0-9]+]] 98 99; GCN: v_cmp_eq_u32_e64 [[CMP0:s\[[0-9]+:[0-9]\]]], s{{[0-9]+}}, v0 100 101; GCN: s_mov_b64 s{{\[}}[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]{{\]}}, exec 102; GCN: s_and_b64 s{{\[}}[[ANDEXEC_LO:[0-9]+]]:[[ANDEXEC_HI:[0-9]+]]{{\]}}, s{{\[}}[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]{{\]}}, [[CMP0]] 103 104; Spill load 105; GCN: buffer_store_dword [[LOAD0]], off, s[0:3], s7 offset:[[LOAD0_OFFSET:[0-9]+]] ; 4-byte Folded Spill 106 107; Spill saved exec 108; VGPR: v_writelane_b32 [[SPILL_VGPR:v[0-9]+]], s[[SAVEEXEC_LO]], [[SAVEEXEC_LO_LANE:[0-9]+]] 109; VGPR: v_writelane_b32 [[SPILL_VGPR]], s[[SAVEEXEC_HI]], [[SAVEEXEC_HI_LANE:[0-9]+]] 110 111 112; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_LO:[0-9]+]], s[[SAVEEXEC_LO]] 113; VMEM: buffer_store_dword v[[V_SAVEEXEC_LO]], off, s[0:3], s7 offset:24 ; 4-byte Folded Spill 114; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_HI:[0-9]+]], s[[SAVEEXEC_HI]] 115; VMEM: buffer_store_dword v[[V_SAVEEXEC_HI]], off, s[0:3], s7 offset:28 ; 4-byte Folded Spill 116 117; GCN: s_mov_b64 exec, s{{\[}}[[ANDEXEC_LO]]:[[ANDEXEC_HI]]{{\]}} 118 119; GCN-NEXT: ; mask branch [[END:BB[0-9]+_[0-9]+]] 120; GCN-NEXT: s_cbranch_execz [[END]] 121 122 123; GCN: [[LOOP:BB[0-9]+_[0-9]+]]: 124; GCN: buffer_load_dword v[[VAL_LOOP_RELOAD:[0-9]+]], off, s[0:3], s7 offset:[[LOAD0_OFFSET]] ; 4-byte Folded Reload 125; GCN: v_subrev_i32_e32 [[VAL_LOOP:v[0-9]+]], vcc, v{{[0-9]+}}, v[[VAL_LOOP_RELOAD]] 126; GCN: v_cmp_ne_u32_e32 vcc, 127; GCN: s_and_b64 vcc, exec, vcc 128; GCN: buffer_store_dword [[VAL_LOOP]], off, s[0:3], s7 offset:[[VAL_SUB_OFFSET:[0-9]+]] ; 4-byte Folded Spill 129; GCN-NEXT: s_cbranch_vccnz [[LOOP]] 130 131 132; GCN: [[END]]: 133; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_LO_LANE]] 134; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_HI_LANE]] 135 136; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_LO:[0-9]+]], off, s[0:3], s7 offset:24 ; 4-byte Folded Reload 137; VMEM: s_waitcnt vmcnt(0) 138; VMEM: v_readfirstlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], v[[V_RELOAD_SAVEEXEC_LO]] 139 140; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_HI:[0-9]+]], off, s[0:3], s7 offset:28 ; 4-byte Folded Reload 141; VMEM: s_waitcnt vmcnt(0) 142; VMEM: v_readfirstlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], v[[V_RELOAD_SAVEEXEC_HI]] 143 144; GCN: s_or_b64 exec, exec, s{{\[}}[[S_RELOAD_SAVEEXEC_LO]]:[[S_RELOAD_SAVEEXEC_HI]]{{\]}} 145; GCN: buffer_load_dword v[[VAL_END:[0-9]+]], off, s[0:3], s7 offset:[[VAL_SUB_OFFSET]] ; 4-byte Folded Reload 146 147; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v[[VAL_END]] 148define amdgpu_kernel void @divergent_loop(i32 addrspace(1)* %out) #0 { 149entry: 150 %tid = call i32 @llvm.amdgcn.workitem.id.x() 151 %load0 = load volatile i32, i32 addrspace(3)* undef 152 %cmp0 = icmp eq i32 %tid, 0 153 br i1 %cmp0, label %loop, label %end 154 155loop: 156 %i = phi i32 [ %i.inc, %loop ], [ 0, %entry ] 157 %val = phi i32 [ %val.sub, %loop ], [ %load0, %entry ] 158 %load1 = load volatile i32, i32 addrspace(3)* undef 159 %i.inc = add i32 %i, 1 160 %val.sub = sub i32 %val, %load1 161 %cmp1 = icmp ne i32 %i, 256 162 br i1 %cmp1, label %loop, label %end 163 164end: 165 %tmp4 = phi i32 [ %val.sub, %loop ], [ 0, %entry ] 166 store i32 %tmp4, i32 addrspace(1)* %out 167 ret void 168} 169 170; GCN-LABEL: {{^}}divergent_if_else_endif: 171; GCN: {{^}}; %bb.0: 172 173; GCN: s_mov_b32 m0, -1 174; GCN: ds_read_b32 [[LOAD0:v[0-9]+]] 175 176; GCN: s_mov_b32 [[ZERO:s[0-9]+]], 0 177; GCN: v_cmp_ne_u32_e64 [[CMP0:s\[[0-9]+:[0-9]\]]], [[ZERO]], v0 178 179; GCN: s_mov_b64 s{{\[}}[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]{{\]}}, exec 180; GCN: s_and_b64 s{{\[}}[[ANDEXEC_LO:[0-9]+]]:[[ANDEXEC_HI:[0-9]+]]{{\]}}, s{{\[}}[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]{{\]}}, [[CMP0]] 181; GCN: s_xor_b64 s{{\[}}[[SAVEEXEC_LO]]:[[SAVEEXEC_HI]]{{\]}}, s{{\[}}[[ANDEXEC_LO]]:[[ANDEXEC_HI]]{{\]}}, s{{\[}}[[SAVEEXEC_LO]]:[[SAVEEXEC_HI]]{{\]}} 182 183; Spill load 184; GCN: buffer_store_dword [[LOAD0]], off, s[0:3], s7 offset:[[LOAD0_OFFSET:[0-9]+]] ; 4-byte Folded Spill 185 186; Spill saved exec 187; VGPR: v_writelane_b32 [[SPILL_VGPR:v[0-9]+]], s[[SAVEEXEC_LO]], [[SAVEEXEC_LO_LANE:[0-9]+]] 188; VGPR: v_writelane_b32 [[SPILL_VGPR]], s[[SAVEEXEC_HI]], [[SAVEEXEC_HI_LANE:[0-9]+]] 189 190; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_LO:[0-9]+]], s[[SAVEEXEC_LO]] 191; VMEM: buffer_store_dword v[[V_SAVEEXEC_LO]], off, s[0:3], s7 offset:[[SAVEEXEC_LO_OFFSET:[0-9]+]] ; 4-byte Folded Spill 192; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_HI:[0-9]+]], s[[SAVEEXEC_HI]] 193; VMEM: buffer_store_dword v[[V_SAVEEXEC_HI]], off, s[0:3], s7 offset:[[SAVEEXEC_HI_OFFSET:[0-9]+]] ; 4-byte Folded Spill 194 195; GCN: s_mov_b64 exec, [[CMP0]] 196 197; FIXME: It makes no sense to put this skip here 198; GCN-NEXT: ; mask branch [[FLOW:BB[0-9]+_[0-9]+]] 199; GCN: s_cbranch_execz [[FLOW]] 200; GCN-NEXT: s_branch [[ELSE:BB[0-9]+_[0-9]+]] 201 202; GCN: [[FLOW]]: ; %Flow 203; VGPR: v_readlane_b32 s[[FLOW_S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_LO_LANE]] 204; VGPR: v_readlane_b32 s[[FLOW_S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_HI_LANE]] 205 206 207; VMEM: buffer_load_dword v[[FLOW_V_RELOAD_SAVEEXEC_LO:[0-9]+]], off, s[0:3], s7 offset:[[SAVEEXEC_LO_OFFSET]] 208; VMEM: s_waitcnt vmcnt(0) 209; VMEM: v_readfirstlane_b32 s[[FLOW_S_RELOAD_SAVEEXEC_LO:[0-9]+]], v[[FLOW_V_RELOAD_SAVEEXEC_LO]] 210 211; VMEM: buffer_load_dword v[[FLOW_V_RELOAD_SAVEEXEC_HI:[0-9]+]], off, s[0:3], s7 offset:[[SAVEEXEC_HI_OFFSET]] ; 4-byte Folded Reload 212; VMEM: s_waitcnt vmcnt(0) 213; VMEM: v_readfirstlane_b32 s[[FLOW_S_RELOAD_SAVEEXEC_HI:[0-9]+]], v[[FLOW_V_RELOAD_SAVEEXEC_HI]] 214 215; GCN: s_or_saveexec_b64 s{{\[}}[[FLOW_S_RELOAD_SAVEEXEC_LO]]:[[FLOW_S_RELOAD_SAVEEXEC_HI]]{{\]}}, s{{\[}}[[FLOW_S_RELOAD_SAVEEXEC_LO]]:[[FLOW_S_RELOAD_SAVEEXEC_HI]]{{\]}} 216 217; Regular spill value restored after exec modification 218; GCN: buffer_load_dword [[FLOW_VAL:v[0-9]+]], off, s[0:3], s7 offset:[[FLOW_VAL_OFFSET:[0-9]+]] ; 4-byte Folded Reload 219 220 221; Spill saved exec 222; VGPR: v_writelane_b32 [[SPILL_VGPR]], s[[FLOW_S_RELOAD_SAVEEXEC_LO]], [[FLOW_SAVEEXEC_LO_LANE:[0-9]+]] 223; VGPR: v_writelane_b32 [[SPILL_VGPR]], s[[FLOW_S_RELOAD_SAVEEXEC_HI]], [[FLOW_SAVEEXEC_HI_LANE:[0-9]+]] 224 225 226; VMEM: v_mov_b32_e32 v[[FLOW_V_SAVEEXEC_LO:[0-9]+]], s[[FLOW_S_RELOAD_SAVEEXEC_LO]] 227; VMEM: buffer_store_dword v[[FLOW_V_SAVEEXEC_LO]], off, s[0:3], s7 offset:[[FLOW_SAVEEXEC_LO_OFFSET:[0-9]+]] ; 4-byte Folded Spill 228; VMEM: v_mov_b32_e32 v[[FLOW_V_SAVEEXEC_HI:[0-9]+]], s[[FLOW_S_RELOAD_SAVEEXEC_HI]] 229; VMEM: buffer_store_dword v[[FLOW_V_SAVEEXEC_HI]], off, s[0:3], s7 offset:[[FLOW_SAVEEXEC_HI_OFFSET:[0-9]+]] ; 4-byte Folded Spill 230 231; GCN: buffer_store_dword [[FLOW_VAL]], off, s[0:3], s7 offset:[[RESULT_OFFSET:[0-9]+]] ; 4-byte Folded Spill 232; GCN: s_xor_b64 exec, exec, s{{\[}}[[FLOW_S_RELOAD_SAVEEXEC_LO]]:[[FLOW_S_RELOAD_SAVEEXEC_HI]]{{\]}} 233; GCN-NEXT: ; mask branch [[ENDIF:BB[0-9]+_[0-9]+]] 234; GCN-NEXT: s_cbranch_execz [[ENDIF]] 235 236 237; GCN: BB{{[0-9]+}}_2: ; %if 238; GCN: ds_read_b32 239; GCN: buffer_load_dword v[[LOAD0_RELOAD:[0-9]+]], off, s[0:3], s7 offset:[[LOAD0_OFFSET]] ; 4-byte Folded Reload 240; GCN: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, v{{[0-9]+}}, v[[LOAD0_RELOAD]] 241; GCN: buffer_store_dword [[ADD]], off, s[0:3], s7 offset:[[RESULT_OFFSET]] ; 4-byte Folded Spill 242; GCN-NEXT: s_branch [[ENDIF:BB[0-9]+_[0-9]+]] 243 244; GCN: [[ELSE]]: ; %else 245; GCN: buffer_load_dword v[[LOAD0_RELOAD:[0-9]+]], off, s[0:3], s7 offset:[[LOAD0_OFFSET]] ; 4-byte Folded Reload 246; GCN: v_subrev_i32_e32 [[SUB:v[0-9]+]], vcc, v{{[0-9]+}}, v[[LOAD0_RELOAD]] 247; GCN: buffer_store_dword [[ADD]], off, s[0:3], s7 offset:[[FLOW_RESULT_OFFSET:[0-9]+]] ; 4-byte Folded Spill 248; GCN-NEXT: s_branch [[FLOW]] 249 250; GCN: [[ENDIF]]: 251; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[FLOW_SAVEEXEC_LO_LANE]] 252; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[FLOW_SAVEEXEC_HI_LANE]] 253 254 255; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_LO:[0-9]+]], off, s[0:3], s7 offset:[[FLOW_SAVEEXEC_LO_OFFSET]] ; 4-byte Folded Reload 256; VMEM: s_waitcnt vmcnt(0) 257; VMEM: v_readfirstlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], v[[V_RELOAD_SAVEEXEC_LO]] 258 259; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_HI:[0-9]+]], off, s[0:3], s7 offset:[[FLOW_SAVEEXEC_HI_OFFSET]] ; 4-byte Folded Reload 260; VMEM: s_waitcnt vmcnt(0) 261; VMEM: v_readfirstlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], v[[V_RELOAD_SAVEEXEC_HI]] 262 263; GCN: s_or_b64 exec, exec, s{{\[}}[[S_RELOAD_SAVEEXEC_LO]]:[[S_RELOAD_SAVEEXEC_HI]]{{\]}} 264 265; GCN: buffer_load_dword v[[RESULT:[0-9]+]], off, s[0:3], s7 offset:[[RESULT_OFFSET]] ; 4-byte Folded Reload 266; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v[[RESULT]] 267define amdgpu_kernel void @divergent_if_else_endif(i32 addrspace(1)* %out) #0 { 268entry: 269 %tid = call i32 @llvm.amdgcn.workitem.id.x() 270 %load0 = load volatile i32, i32 addrspace(3)* undef 271 %cmp0 = icmp eq i32 %tid, 0 272 br i1 %cmp0, label %if, label %else 273 274if: 275 %load1 = load volatile i32, i32 addrspace(3)* undef 276 %val0 = add i32 %load0, %load1 277 br label %endif 278 279else: 280 %load2 = load volatile i32, i32 addrspace(3)* undef 281 %val1 = sub i32 %load0, %load2 282 br label %endif 283 284endif: 285 %result = phi i32 [ %val0, %if ], [ %val1, %else ] 286 store i32 %result, i32 addrspace(1)* %out 287 ret void 288} 289 290declare i32 @llvm.amdgcn.workitem.id.x() #1 291 292attributes #0 = { nounwind } 293attributes #1 = { nounwind readnone } 294