llvm/test/CodeGen/AMDGPU/lower-work-group-id-intrinsics-pal.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=-architected-sgprs -global-isel=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-SDAG %s
   3 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=-architected-sgprs -global-isel=1 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-GISEL %s
   4 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=+architected-sgprs -global-isel=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9ARCH-SDAG %s
   5 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=+architected-sgprs -global-isel=1 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9ARCH-GISEL %s
   6 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -global-isel=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG %s
   7 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -global-isel=1 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL %s
   8
   9 define amdgpu_cs void @_amdgpu_cs_main() {
  10 ; GFX9-LABEL: _amdgpu_cs_main:
  11 ; GFX9:       ; %bb.0: ; %.entry
  12 ; GFX9-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  13 ; GFX9-NEXT:    s_endpgm
  14 ;
  15 ; GFX9-SDAG-LABEL: _amdgpu_cs_main:
  16 ; GFX9-SDAG:       ; %bb.0: ; %.entry
  17 ; GFX9-SDAG-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  18 ; GFX9-SDAG-NEXT:    s_endpgm
  19 ;
  20 ; GFX9-GISEL-LABEL: _amdgpu_cs_main:
  21 ; GFX9-GISEL:       ; %bb.0: ; %.entry
  22 ; GFX9-GISEL-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  23 ; GFX9-GISEL-NEXT:    s_endpgm
  24 ;
  25 ; GFX9ARCH-SDAG-LABEL: _amdgpu_cs_main:
  26 ; GFX9ARCH-SDAG:       ; %bb.0: ; %.entry
  27 ; GFX9ARCH-SDAG-NEXT:    s_lshr_b32 s0, ttmp7, 16
  28 ; GFX9ARCH-SDAG-NEXT:    s_and_b32 s1, ttmp7, 0xffff
  29 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, ttmp9
  30 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v1, s1
  31 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v2, s0
  32 ; GFX9ARCH-SDAG-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  33 ; GFX9ARCH-SDAG-NEXT:    s_endpgm
  34 ;
  35 ; GFX9ARCH-GISEL-LABEL: _amdgpu_cs_main:
  36 ; GFX9ARCH-GISEL:       ; %bb.0: ; %.entry
  37 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s0, ttmp9
  38 ; GFX9ARCH-GISEL-NEXT:    s_and_b32 s1, ttmp7, 0xffff
  39 ; GFX9ARCH-GISEL-NEXT:    s_lshr_b32 s2, ttmp7, 16
  40 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, s0
  41 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v1, s1
  42 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v2, s2
  43 ; GFX9ARCH-GISEL-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  44 ; GFX9ARCH-GISEL-NEXT:    s_endpgm
  45 ;
  46 ; GFX12-SDAG-LABEL: _amdgpu_cs_main:
  47 ; GFX12-SDAG:       ; %bb.0: ; %.entry
  48 ; GFX12-SDAG-NEXT:    s_and_b32 s0, ttmp7, 0xffff
  49 ; GFX12-SDAG-NEXT:    s_lshr_b32 s1, ttmp7, 16
  50 ; GFX12-SDAG-NEXT:    v_dual_mov_b32 v0, ttmp9 :: v_dual_mov_b32 v1, s0
  51 ; GFX12-SDAG-NEXT:    v_mov_b32_e32 v2, s1
  52 ; GFX12-SDAG-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], null
  53 ; GFX12-SDAG-NEXT:    s_endpgm
  54 ;
  55 ; GFX12-GISEL-LABEL: _amdgpu_cs_main:
  56 ; GFX12-GISEL:       ; %bb.0: ; %.entry
  57 ; GFX12-GISEL-NEXT:    s_mov_b32 s0, ttmp9
  58 ; GFX12-GISEL-NEXT:    s_and_b32 s1, ttmp7, 0xffff
  59 ; GFX12-GISEL-NEXT:    s_lshr_b32 s2, ttmp7, 16
  60 ; GFX12-GISEL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
  61 ; GFX12-GISEL-NEXT:    v_mov_b32_e32 v2, s2
  62 ; GFX12-GISEL-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], null
  63 ; GFX12-GISEL-NEXT:    s_endpgm
  64 .entry:
  65   %idx = call i32 @llvm.amdgcn.workgroup.id.x()
  66   %idy = call i32 @llvm.amdgcn.workgroup.id.y()
  67   %idz = call i32 @llvm.amdgcn.workgroup.id.z()
  68   %ielemx = insertelement <3 x i32> undef, i32 %idx, i64 0
  69   %ielemy = insertelement <3 x i32> %ielemx, i32 %idy, i64 1
  70   %ielemz = insertelement <3 x i32> %ielemy, i32 %idz, i64 2
  71   call void @llvm.amdgcn.raw.ptr.buffer.store.v3i32(<3 x i32> %ielemz, ptr addrspace(8) undef, i32 0, i32 0, i32 0)
  72   ret void
  73 }
  74
  75 define amdgpu_cs void @caller() {
  76 ; GFX9-SDAG-LABEL: caller:
  77 ; GFX9-SDAG:       ; %bb.0:
  78 ; GFX9-SDAG-NEXT:    s_getpc_b64 s[8:9]
  79 ; GFX9-SDAG-NEXT:    s_mov_b32 s8, s0
  80 ; GFX9-SDAG-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x10
  81 ; GFX9-SDAG-NEXT:    s_mov_b32 s5, callee@abs32@hi
  82 ; GFX9-SDAG-NEXT:    s_mov_b32 s4, callee@abs32@lo
  83 ; GFX9-SDAG-NEXT:    s_mov_b32 s32, 0
  84 ; GFX9-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  85 ; GFX9-SDAG-NEXT:    s_add_u32 s8, s8, s0
  86 ; GFX9-SDAG-NEXT:    s_addc_u32 s9, s9, 0
  87 ; GFX9-SDAG-NEXT:    s_mov_b64 s[0:1], s[8:9]
  88 ; GFX9-SDAG-NEXT:    s_mov_b64 s[2:3], s[10:11]
  89 ; GFX9-SDAG-NEXT:    s_swappc_b64 s[30:31], s[4:5]
  90 ; GFX9-SDAG-NEXT:    s_endpgm
  91 ;
  92 ; GFX9-GISEL-LABEL: caller:
  93 ; GFX9-GISEL:       ; %bb.0:
  94 ; GFX9-GISEL-NEXT:    s_getpc_b64 s[8:9]
  95 ; GFX9-GISEL-NEXT:    s_mov_b32 s8, s0
  96 ; GFX9-GISEL-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x10
  97 ; GFX9-GISEL-NEXT:    s_mov_b32 s4, callee@abs32@lo
  98 ; GFX9-GISEL-NEXT:    s_mov_b32 s5, callee@abs32@hi
  99 ; GFX9-GISEL-NEXT:    s_mov_b32 s32, 0
 100 ; GFX9-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 101 ; GFX9-GISEL-NEXT:    s_add_u32 s8, s8, s0
 102 ; GFX9-GISEL-NEXT:    s_addc_u32 s9, s9, 0
 103 ; GFX9-GISEL-NEXT:    s_mov_b64 s[0:1], s[8:9]
 104 ; GFX9-GISEL-NEXT:    s_mov_b64 s[2:3], s[10:11]
 105 ; GFX9-GISEL-NEXT:    s_swappc_b64 s[30:31], s[4:5]
 106 ; GFX9-GISEL-NEXT:    s_endpgm
 107 ;
 108 ; GFX9ARCH-SDAG-LABEL: caller:
 109 ; GFX9ARCH-SDAG:       ; %bb.0:
 110 ; GFX9ARCH-SDAG-NEXT:    s_getpc_b64 s[8:9]
 111 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s8, s0
 112 ; GFX9ARCH-SDAG-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x10
 113 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s5, callee@abs32@hi
 114 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s4, callee@abs32@lo
 115 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, ttmp9
 116 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s32, 0
 117 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 118 ; GFX9ARCH-SDAG-NEXT:    s_add_u32 s8, s8, s0
 119 ; GFX9ARCH-SDAG-NEXT:    s_addc_u32 s9, s9, 0
 120 ; GFX9ARCH-SDAG-NEXT:    s_mov_b64 s[0:1], s[8:9]
 121 ; GFX9ARCH-SDAG-NEXT:    s_mov_b64 s[2:3], s[10:11]
 122 ; GFX9ARCH-SDAG-NEXT:    s_swappc_b64 s[30:31], s[4:5]
 123 ; GFX9ARCH-SDAG-NEXT:    s_endpgm
 124 ;
 125 ; GFX9ARCH-GISEL-LABEL: caller:
 126 ; GFX9ARCH-GISEL:       ; %bb.0:
 127 ; GFX9ARCH-GISEL-NEXT:    s_getpc_b64 s[8:9]
 128 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s8, s0
 129 ; GFX9ARCH-GISEL-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x10
 130 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s4, callee@abs32@lo
 131 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s5, callee@abs32@hi
 132 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, ttmp9
 133 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s32, 0
 134 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 135 ; GFX9ARCH-GISEL-NEXT:    s_add_u32 s8, s8, s0
 136 ; GFX9ARCH-GISEL-NEXT:    s_addc_u32 s9, s9, 0
 137 ; GFX9ARCH-GISEL-NEXT:    s_mov_b64 s[0:1], s[8:9]
 138 ; GFX9ARCH-GISEL-NEXT:    s_mov_b64 s[2:3], s[10:11]
 139 ; GFX9ARCH-GISEL-NEXT:    s_swappc_b64 s[30:31], s[4:5]
 140 ; GFX9ARCH-GISEL-NEXT:    s_endpgm
 141 ;
 142 ; GFX12-SDAG-LABEL: caller:
 143 ; GFX12-SDAG:       ; %bb.0:
 144 ; GFX12-SDAG-NEXT:    v_mov_b32_e32 v0, ttmp9
 145 ; GFX12-SDAG-NEXT:    s_mov_b32 s1, callee@abs32@hi
 146 ; GFX12-SDAG-NEXT:    s_mov_b32 s0, callee@abs32@lo
 147 ; GFX12-SDAG-NEXT:    s_mov_b32 s32, 0
 148 ; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
 149 ; GFX12-SDAG-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 150 ; GFX12-SDAG-NEXT:    s_endpgm
 151 ;
 152 ; GFX12-GISEL-LABEL: caller:
 153 ; GFX12-GISEL:       ; %bb.0:
 154 ; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, ttmp9
 155 ; GFX12-GISEL-NEXT:    s_mov_b32 s0, callee@abs32@lo
 156 ; GFX12-GISEL-NEXT:    s_mov_b32 s1, callee@abs32@hi
 157 ; GFX12-GISEL-NEXT:    s_mov_b32 s32, 0
 158 ; GFX12-GISEL-NEXT:    s_wait_alu 0xfffe
 159 ; GFX12-GISEL-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 160 ; GFX12-GISEL-NEXT:    s_endpgm
 161   %idx = call i32 @llvm.amdgcn.workgroup.id.x()
 162   call amdgpu_gfx void @callee(i32 %idx)
 163   ret void
 164 }
 165
 166 declare amdgpu_gfx void @callee(i32)
 167
 168 define amdgpu_gfx void @workgroup_ids_gfx(ptr addrspace(1) %outx, ptr addrspace(1) %outy, ptr addrspace(1) %outz) {
 169 ; GFX9-SDAG-LABEL: workgroup_ids_gfx:
 170 ; GFX9-SDAG:       ; %bb.0:
 171 ; GFX9-SDAG-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 172 ; GFX9-SDAG-NEXT:    global_store_dword v[0:1], v0, off
 173 ; GFX9-SDAG-NEXT:    s_waitcnt vmcnt(0)
 174 ; GFX9-SDAG-NEXT:    global_store_dword v[2:3], v0, off
 175 ; GFX9-SDAG-NEXT:    s_waitcnt vmcnt(0)
 176 ; GFX9-SDAG-NEXT:    global_store_dword v[4:5], v0, off
 177 ; GFX9-SDAG-NEXT:    s_waitcnt vmcnt(0)
 178 ; GFX9-SDAG-NEXT:    s_setpc_b64 s[30:31]
 179 ;
 180 ; GFX9-GISEL-LABEL: workgroup_ids_gfx:
 181 ; GFX9-GISEL:       ; %bb.0:
 182 ; GFX9-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 183 ; GFX9-GISEL-NEXT:    s_setpc_b64 s[30:31]
 184 ;
 185 ; GFX9ARCH-SDAG-LABEL: workgroup_ids_gfx:
 186 ; GFX9ARCH-SDAG:       ; %bb.0:
 187 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 188 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v6, ttmp9
 189 ; GFX9ARCH-SDAG-NEXT:    s_and_b32 s34, ttmp7, 0xffff
 190 ; GFX9ARCH-SDAG-NEXT:    global_store_dword v[0:1], v6, off
 191 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0)
 192 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, s34
 193 ; GFX9ARCH-SDAG-NEXT:    s_lshr_b32 s34, ttmp7, 16
 194 ; GFX9ARCH-SDAG-NEXT:    global_store_dword v[2:3], v0, off
 195 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0)
 196 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, s34
 197 ; GFX9ARCH-SDAG-NEXT:    global_store_dword v[4:5], v0, off
 198 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0)
 199 ; GFX9ARCH-SDAG-NEXT:    s_setpc_b64 s[30:31]
 200 ;
 201 ; GFX9ARCH-GISEL-LABEL: workgroup_ids_gfx:
 202 ; GFX9ARCH-GISEL:       ; %bb.0:
 203 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 204 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v6, ttmp9
 205 ; GFX9ARCH-GISEL-NEXT:    s_and_b32 s34, ttmp7, 0xffff
 206 ; GFX9ARCH-GISEL-NEXT:    s_lshr_b32 s35, ttmp7, 16
 207 ; GFX9ARCH-GISEL-NEXT:    global_store_dword v[0:1], v6, off
 208 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0)
 209 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, s34
 210 ; GFX9ARCH-GISEL-NEXT:    global_store_dword v[2:3], v0, off
 211 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0)
 212 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, s35
 213 ; GFX9ARCH-GISEL-NEXT:    global_store_dword v[4:5], v0, off
 214 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0)
 215 ; GFX9ARCH-GISEL-NEXT:    s_setpc_b64 s[30:31]
 216 ;
 217 ; GFX12-LABEL: workgroup_ids_gfx:
 218 ; GFX12:       ; %bb.0:
 219 ; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
 220 ; GFX12-NEXT:    s_wait_expcnt 0x0
 221 ; GFX12-NEXT:    s_wait_samplecnt 0x0
 222 ; GFX12-NEXT:    s_wait_bvhcnt 0x0
 223 ; GFX12-NEXT:    s_wait_kmcnt 0x0
 224 ; GFX12-NEXT:    s_and_b32 s0, ttmp7, 0xffff
 225 ; GFX12-NEXT:    s_wait_alu 0xfffe
 226 ; GFX12-NEXT:    v_dual_mov_b32 v6, ttmp9 :: v_dual_mov_b32 v7, s0
 227 ; GFX12-NEXT:    s_lshr_b32 s1, ttmp7, 16
 228 ; GFX12-NEXT:    s_wait_alu 0xfffe
 229 ; GFX12-NEXT:    v_mov_b32_e32 v8, s1
 230 ; GFX12-NEXT:    s_wait_storecnt 0x0
 231 ; GFX12-NEXT:    global_store_b32 v[0:1], v6, off scope:SCOPE_SYS
 232 ; GFX12-NEXT:    s_wait_storecnt 0x0
 233 ; GFX12-NEXT:    global_store_b32 v[2:3], v7, off scope:SCOPE_SYS
 234 ; GFX12-NEXT:    s_wait_storecnt 0x0
 235 ; GFX12-NEXT:    global_store_b32 v[4:5], v8, off scope:SCOPE_SYS
 236 ; GFX12-NEXT:    s_wait_storecnt 0x0
 237 ; GFX12-NEXT:    s_setpc_b64 s[30:31]
 238   %id.x = call i32 @llvm.amdgcn.workgroup.id.x()
 239   %id.y = call i32 @llvm.amdgcn.workgroup.id.y()
 240   %id.z = call i32 @llvm.amdgcn.workgroup.id.z()
 241   store volatile i32 %id.x, ptr addrspace(1) %outx
 242   store volatile i32 %id.y, ptr addrspace(1) %outy
 243   store volatile i32 %id.z, ptr addrspace(1) %outz
 244   ret void
 245 }