llvm/test/CodeGen/AMDGPU/llvm.amdgcn.workgroup.id.ll

   1 ; RUN: llc -march=amdgcn -mtriple=amdgcn-unknown-amdhsa --amdhsa-code-object-version=2 -mcpu=kaveri -verify-machineinstrs < %s | FileCheck --check-prefixes=ALL,CO-V2  %s
   2 ; RUN: llc -march=amdgcn -mtriple=amdgcn-unknown-amdhsa --amdhsa-code-object-version=2 -mcpu=carrizo -verify-machineinstrs < %s | FileCheck --check-prefixes=ALL,CO-V2  %s
   3 ; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck --check-prefixes=ALL,UNKNOWN-OS %s
   4 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck --check-prefixes=ALL,UNKNOWN-OS %s
   5 ; RUN: llc -mtriple=amdgcn-unknown-mesa3d -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefixes=ALL,CO-V2 %s
   6 ; RUN: llc -mtriple=amdgcn-unknown-mesa3d -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefixes=ALL,CO-V2 %s
   7
   8 declare i32 @llvm.amdgcn.workgroup.id.x() #0
   9 declare i32 @llvm.amdgcn.workgroup.id.y() #0
  10 declare i32 @llvm.amdgcn.workgroup.id.z() #0
  11
  12 ; ALL-LABEL: {{^}}test_workgroup_id_x:
  13
  14 ; CO-V2: .amd_kernel_code_t
  15 ; CO-V2: user_sgpr_count = 6
  16 ; CO-V2: enable_sgpr_workgroup_id_x = 1
  17 ; CO-V2: enable_sgpr_workgroup_id_y = 0
  18 ; CO-V2: enable_sgpr_workgroup_id_z = 0
  19 ; CO-V2: enable_sgpr_workgroup_info = 0
  20 ; CO-V2: enable_vgpr_workitem_id = 0
  21 ; CO-V2: enable_sgpr_grid_workgroup_count_x = 0
  22 ; CO-V2: enable_sgpr_grid_workgroup_count_y = 0
  23 ; CO-V2: enable_sgpr_grid_workgroup_count_z = 0
  24 ; CO-V2: .end_amd_kernel_code_t
  25
  26 ; UNKNOWN-OS: v_mov_b32_e32 [[VCOPY:v[0-9]+]], s2{{$}}
  27 ; CO-V2: v_mov_b32_e32 [[VCOPY:v[0-9]+]], s6{{$}}
  28
  29 ; ALL: {{buffer|flat}}_store_dword {{.*}}[[VCOPY]]
  30
  31 ; CO-V2: COMPUTE_PGM_RSRC2:USER_SGPR: 6
  32 ; ALL-NOCO-V2: COMPUTE_PGM_RSRC2:USER_SGPR: 2
  33 ; ALL: COMPUTE_PGM_RSRC2:TGID_X_EN: 1
  34 ; ALL: COMPUTE_PGM_RSRC2:TGID_Y_EN: 0
  35 ; ALL: COMPUTE_PGM_RSRC2:TGID_Z_EN: 0
  36 ; ALL: COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: 0
  37 define amdgpu_kernel void @test_workgroup_id_x(i32 addrspace(1)* %out) #1 {
  38   %id = call i32 @llvm.amdgcn.workgroup.id.x()
  39   store i32 %id, i32 addrspace(1)* %out
  40   ret void
  41 }
  42
  43 ; ALL-LABEL: {{^}}test_workgroup_id_y:
  44 ; CO-V2: user_sgpr_count = 6
  45 ; CO-V2: enable_sgpr_workgroup_id_x = 1
  46 ; CO-V2: enable_sgpr_workgroup_id_y = 1
  47 ; CO-V2: enable_sgpr_workgroup_id_z = 0
  48 ; CO-V2: enable_sgpr_workgroup_info = 0
  49 ; CO-V2: enable_sgpr_grid_workgroup_count_x = 0
  50 ; CO-V2: enable_sgpr_grid_workgroup_count_y = 0
  51 ; CO-V2: enable_sgpr_grid_workgroup_count_z = 0
  52
  53 ; UNKNOWN-OS: v_mov_b32_e32 [[VCOPY:v[0-9]+]], s3{{$}}
  54 ; HSA: v_mov_b32_e32 [[VCOPY:v[0-9]+]], s7{{$}}
  55
  56 ; ALL: {{buffer|flat}}_store_dword {{.*}}[[VCOPY]]
  57
  58 ; CO-V2: COMPUTE_PGM_RSRC2:USER_SGPR: 6
  59 ; ALL-NOCO-V2: COMPUTE_PGM_RSRC2:USER_SGPR: 2
  60 ; ALL: COMPUTE_PGM_RSRC2:TGID_X_EN: 1
  61 ; ALL: COMPUTE_PGM_RSRC2:TGID_Y_EN: 1
  62 ; ALL: COMPUTE_PGM_RSRC2:TGID_Z_EN: 0
  63 ; ALL: COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: 0
  64 define amdgpu_kernel void @test_workgroup_id_y(i32 addrspace(1)* %out) #1 {
  65   %id = call i32 @llvm.amdgcn.workgroup.id.y()
  66   store i32 %id, i32 addrspace(1)* %out
  67   ret void
  68 }
  69
  70 ; ALL-LABEL: {{^}}test_workgroup_id_z:
  71 ; CO-V2: user_sgpr_count = 6
  72 ; CO-V2: enable_sgpr_workgroup_id_x = 1
  73 ; CO-V2: enable_sgpr_workgroup_id_y = 0
  74 ; CO-V2: enable_sgpr_workgroup_id_z = 1
  75 ; CO-V2: enable_sgpr_workgroup_info = 0
  76 ; CO-V2: enable_vgpr_workitem_id = 0
  77 ; CO-V2: enable_sgpr_private_segment_buffer = 1
  78 ; CO-V2: enable_sgpr_dispatch_ptr = 0
  79 ; CO-V2: enable_sgpr_queue_ptr = 0
  80 ; CO-V2: enable_sgpr_kernarg_segment_ptr = 1
  81 ; CO-V2: enable_sgpr_dispatch_id = 0
  82 ; CO-V2: enable_sgpr_flat_scratch_init = 0
  83 ; CO-V2: enable_sgpr_private_segment_size = 0
  84 ; CO-V2: enable_sgpr_grid_workgroup_count_x = 0
  85 ; CO-V2: enable_sgpr_grid_workgroup_count_y = 0
  86 ; CO-V2: enable_sgpr_grid_workgroup_count_z = 0
  87
  88 ; UNKNOWN-OS: v_mov_b32_e32 [[VCOPY:v[0-9]+]], s3{{$}}
  89 ; HSA: v_mov_b32_e32 [[VCOPY:v[0-9]+]], s7{{$}}
  90
  91 ; ALL: {{buffer|flat}}_store_dword {{.*}}[[VCOPY]]
  92
  93 ; CO-V2: COMPUTE_PGM_RSRC2:USER_SGPR: 6
  94 ; ALL-NOCO-V2: COMPUTE_PGM_RSRC2:USER_SGPR: 2
  95 ; ALL: COMPUTE_PGM_RSRC2:TGID_X_EN: 1
  96 ; ALL: COMPUTE_PGM_RSRC2:TGID_Y_EN: 0
  97 ; ALL: COMPUTE_PGM_RSRC2:TGID_Z_EN: 1
  98 ; ALL: COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: 0
  99 define amdgpu_kernel void @test_workgroup_id_z(i32 addrspace(1)* %out) #1 {
 100   %id = call i32 @llvm.amdgcn.workgroup.id.z()
 101   store i32 %id, i32 addrspace(1)* %out
 102   ret void
 103 }
 104
 105 attributes #0 = { nounwind readnone }
 106 attributes #1 = { nounwind }