llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.tbuffer.store.d16.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc < %s -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck -check-prefixes=PREGFX10-UNPACKED %s
   3 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx810 -verify-machineinstrs | FileCheck -check-prefixes=PREGFX10-PACKED %s
   4 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs | FileCheck -check-prefixes=PREGFX10-PACKED %s
   5 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -check-prefixes=GFX10-PACKED %s
   6 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs | FileCheck -check-prefixes=GFX11-PACKED %s
   7 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -amdgpu-enable-vopd=0 -verify-machineinstrs | FileCheck -check-prefixes=GFX12-PACKED,GFX12-PACKED-SDAG %s
   8 ; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1200 -amdgpu-enable-vopd=0 -verify-machineinstrs | FileCheck -check-prefixes=GFX12-PACKED,GFX12-PACKED-GISEL %s
   9
  10 define amdgpu_kernel void @tbuffer_store_d16_x(<4 x i32> %rsrc, half %data) {
  11 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_x:
  12 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
  13 ; PREGFX10-UNPACKED-NEXT:    s_load_dword s4, s[2:3], 0x34
  14 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
  15 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
  16 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
  17 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_x v0, off, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED]
  18 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
  19 ;
  20 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_x:
  21 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
  22 ; PREGFX10-PACKED-NEXT:    s_load_dword s0, s[2:3], 0x34
  23 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
  24 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  25 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
  26 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_x v0, off, s[4:7], 0 format:[BUF_NUM_FORMAT_USCALED]
  27 ; PREGFX10-PACKED-NEXT:    s_endpgm
  28 ;
  29 ; GFX10-PACKED-LABEL: tbuffer_store_d16_x:
  30 ; GFX10-PACKED:       ; %bb.0: ; %main_body
  31 ; GFX10-PACKED-NEXT:    s_clause 0x1
  32 ; GFX10-PACKED-NEXT:    s_load_dword s0, s[2:3], 0x34
  33 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
  34 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  35 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
  36 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_x v0, off, s[4:7], 0 format:[BUF_FMT_10_11_11_SSCALED]
  37 ; GFX10-PACKED-NEXT:    s_endpgm
  38 ;
  39 ; GFX11-PACKED-LABEL: tbuffer_store_d16_x:
  40 ; GFX11-PACKED:       ; %bb.0: ; %main_body
  41 ; GFX11-PACKED-NEXT:    s_clause 0x1
  42 ; GFX11-PACKED-NEXT:    s_load_b32 s4, s[2:3], 0x34
  43 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
  44 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  45 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
  46 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_x v0, off, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM]
  47 ; GFX11-PACKED-NEXT:    s_nop 0
  48 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  49 ; GFX11-PACKED-NEXT:    s_endpgm
  50 ;
  51 ; GFX12-PACKED-LABEL: tbuffer_store_d16_x:
  52 ; GFX12-PACKED:       ; %bb.0: ; %main_body
  53 ; GFX12-PACKED-NEXT:    s_clause 0x1
  54 ; GFX12-PACKED-NEXT:    s_load_b32 s4, s[2:3], 0x34
  55 ; GFX12-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
  56 ; GFX12-PACKED-NEXT:    s_wait_kmcnt 0x0
  57 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, s4
  58 ; GFX12-PACKED-NEXT:    tbuffer_store_d16_format_x v0, off, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM]
  59 ; GFX12-PACKED-NEXT:    s_nop 0
  60 ; GFX12-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  61 ; GFX12-PACKED-NEXT:    s_endpgm
  62 main_body:
  63   call void @llvm.amdgcn.raw.tbuffer.store.f16(half %data, <4 x i32> %rsrc, i32 0, i32 0, i32 33, i32 0)
  64   ret void
  65 }
  66
  67 define amdgpu_kernel void @tbuffer_store_d16_xy(<4 x i32> %rsrc, <2 x half> %data) {
  68 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xy:
  69 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
  70 ; PREGFX10-UNPACKED-NEXT:    s_load_dword s4, s[2:3], 0x34
  71 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
  72 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
  73 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s5, s4, 16
  74 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s4, s4, 0xffff
  75 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
  76 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v1, s5
  77 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_xy v[0:1], off, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED]
  78 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
  79 ;
  80 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xy:
  81 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
  82 ; PREGFX10-PACKED-NEXT:    s_load_dword s0, s[2:3], 0x34
  83 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
  84 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  85 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
  86 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_xy v0, off, s[4:7], 0 format:[BUF_NUM_FORMAT_USCALED]
  87 ; PREGFX10-PACKED-NEXT:    s_endpgm
  88 ;
  89 ; GFX10-PACKED-LABEL: tbuffer_store_d16_xy:
  90 ; GFX10-PACKED:       ; %bb.0: ; %main_body
  91 ; GFX10-PACKED-NEXT:    s_clause 0x1
  92 ; GFX10-PACKED-NEXT:    s_load_dword s0, s[2:3], 0x34
  93 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
  94 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  95 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
  96 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_xy v0, off, s[4:7], 0 format:[BUF_FMT_10_11_11_SSCALED]
  97 ; GFX10-PACKED-NEXT:    s_endpgm
  98 ;
  99 ; GFX11-PACKED-LABEL: tbuffer_store_d16_xy:
 100 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 101 ; GFX11-PACKED-NEXT:    s_clause 0x1
 102 ; GFX11-PACKED-NEXT:    s_load_b32 s4, s[2:3], 0x34
 103 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
 104 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 105 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 106 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_xy v0, off, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM]
 107 ; GFX11-PACKED-NEXT:    s_nop 0
 108 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 109 ; GFX11-PACKED-NEXT:    s_endpgm
 110 ;
 111 ; GFX12-PACKED-LABEL: tbuffer_store_d16_xy:
 112 ; GFX12-PACKED:       ; %bb.0: ; %main_body
 113 ; GFX12-PACKED-NEXT:    s_clause 0x1
 114 ; GFX12-PACKED-NEXT:    s_load_b32 s4, s[2:3], 0x34
 115 ; GFX12-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
 116 ; GFX12-PACKED-NEXT:    s_wait_kmcnt 0x0
 117 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 118 ; GFX12-PACKED-NEXT:    tbuffer_store_d16_format_xy v0, off, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM]
 119 ; GFX12-PACKED-NEXT:    s_nop 0
 120 ; GFX12-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 121 ; GFX12-PACKED-NEXT:    s_endpgm
 122 main_body:
 123   call void @llvm.amdgcn.raw.tbuffer.store.v2f16(<2 x half> %data, <4 x i32> %rsrc, i32 0, i32 0, i32 33, i32 0)
 124   ret void
 125 }
 126
 127 define amdgpu_kernel void @tbuffer_store_d16_xyz(<4 x i32> %rsrc, <4 x half> %data) {
 128 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xyz:
 129 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
 130 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x34
 131 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
 132 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
 133 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 134 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s6, s4, 16
 135 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s4, s4, 0xffff
 136 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
 137 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v1, s6
 138 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v2, s5
 139 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_xyz v[0:2], off, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED]
 140 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
 141 ;
 142 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xyz:
 143 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
 144 ; PREGFX10-PACKED-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 145 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 146 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 147 ; PREGFX10-PACKED-NEXT:    s_and_b32 s1, s1, 0xffff
 148 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
 149 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s1
 150 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyz v[0:1], off, s[4:7], 0 format:[BUF_NUM_FORMAT_USCALED]
 151 ; PREGFX10-PACKED-NEXT:    s_endpgm
 152 ;
 153 ; GFX10-PACKED-LABEL: tbuffer_store_d16_xyz:
 154 ; GFX10-PACKED:       ; %bb.0: ; %main_body
 155 ; GFX10-PACKED-NEXT:    s_clause 0x1
 156 ; GFX10-PACKED-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 157 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 158 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 159 ; GFX10-PACKED-NEXT:    s_and_b32 s1, s1, 0xffff
 160 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
 161 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s1
 162 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyz v[0:1], off, s[4:7], 0 format:[BUF_FMT_10_11_11_SSCALED]
 163 ; GFX10-PACKED-NEXT:    s_endpgm
 164 ;
 165 ; GFX11-PACKED-LABEL: tbuffer_store_d16_xyz:
 166 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 167 ; GFX11-PACKED-NEXT:    s_clause 0x1
 168 ; GFX11-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x34
 169 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
 170 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 171 ; GFX11-PACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 172 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 173 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 174 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM]
 175 ; GFX11-PACKED-NEXT:    s_nop 0
 176 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 177 ; GFX11-PACKED-NEXT:    s_endpgm
 178 ;
 179 ; GFX12-PACKED-SDAG-LABEL: tbuffer_store_d16_xyz:
 180 ; GFX12-PACKED-SDAG:       ; %bb.0: ; %main_body
 181 ; GFX12-PACKED-SDAG-NEXT:    s_clause 0x1
 182 ; GFX12-PACKED-SDAG-NEXT:    s_load_b64 s[4:5], s[2:3], 0x34
 183 ; GFX12-PACKED-SDAG-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
 184 ; GFX12-PACKED-SDAG-NEXT:    s_wait_kmcnt 0x0
 185 ; GFX12-PACKED-SDAG-NEXT:    s_and_b32 s5, s5, 0xffff
 186 ; GFX12-PACKED-SDAG-NEXT:    v_mov_b32_e32 v0, s4
 187 ; GFX12-PACKED-SDAG-NEXT:    v_mov_b32_e32 v1, s5
 188 ; GFX12-PACKED-SDAG-NEXT:    tbuffer_store_d16_format_xyz v[0:1], off, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM]
 189 ; GFX12-PACKED-SDAG-NEXT:    s_nop 0
 190 ; GFX12-PACKED-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 191 ; GFX12-PACKED-SDAG-NEXT:    s_endpgm
 192 ;
 193 ; GFX12-PACKED-GISEL-LABEL: tbuffer_store_d16_xyz:
 194 ; GFX12-PACKED-GISEL:       ; %bb.0: ; %main_body
 195 ; GFX12-PACKED-GISEL-NEXT:    s_clause 0x1
 196 ; GFX12-PACKED-GISEL-NEXT:    s_load_b64 s[4:5], s[2:3], 0x34
 197 ; GFX12-PACKED-GISEL-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
 198 ; GFX12-PACKED-GISEL-NEXT:    s_wait_kmcnt 0x0
 199 ; GFX12-PACKED-GISEL-NEXT:    s_pack_lh_b32_b16 s4, s4, s4
 200 ; GFX12-PACKED-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 201 ; GFX12-PACKED-GISEL-NEXT:    v_mov_b32_e32 v0, s4
 202 ; GFX12-PACKED-GISEL-NEXT:    v_mov_b32_e32 v1, s5
 203 ; GFX12-PACKED-GISEL-NEXT:    tbuffer_store_d16_format_xyzw v[0:1], off, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM]
 204 ; GFX12-PACKED-GISEL-NEXT:    s_nop 0
 205 ; GFX12-PACKED-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 206 ; GFX12-PACKED-GISEL-NEXT:    s_endpgm
 207 main_body:
 208   %data_subvec = shufflevector <4 x half> %data, <4 x half> undef, <3 x i32> <i32 0, i32 1, i32 2>
 209   call void @llvm.amdgcn.raw.tbuffer.store.v3f16(<3 x half> %data_subvec, <4 x i32> %rsrc, i32 0, i32 0, i32 33, i32 0)
 210   ret void
 211 }
 212
 213 define amdgpu_kernel void @tbuffer_store_d16_xyzw(<4 x i32> %rsrc, <4 x half> %data) {
 214 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xyzw:
 215 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
 216 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x34
 217 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
 218 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
 219 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s6, s5, 16
 220 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 221 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s7, s4, 16
 222 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s4, s4, 0xffff
 223 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
 224 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v1, s7
 225 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v2, s5
 226 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v3, s6
 227 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_xyzw v[0:3], off, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED]
 228 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
 229 ;
 230 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xyzw:
 231 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
 232 ; PREGFX10-PACKED-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 233 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 234 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 235 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
 236 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s1
 237 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyzw v[0:1], off, s[4:7], 0 format:[BUF_NUM_FORMAT_USCALED]
 238 ; PREGFX10-PACKED-NEXT:    s_endpgm
 239 ;
 240 ; GFX10-PACKED-LABEL: tbuffer_store_d16_xyzw:
 241 ; GFX10-PACKED:       ; %bb.0: ; %main_body
 242 ; GFX10-PACKED-NEXT:    s_clause 0x1
 243 ; GFX10-PACKED-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 244 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 245 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 246 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s0
 247 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s1
 248 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyzw v[0:1], off, s[4:7], 0 format:[BUF_FMT_10_11_11_SSCALED]
 249 ; GFX10-PACKED-NEXT:    s_endpgm
 250 ;
 251 ; GFX11-PACKED-LABEL: tbuffer_store_d16_xyzw:
 252 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 253 ; GFX11-PACKED-NEXT:    s_clause 0x1
 254 ; GFX11-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x34
 255 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
 256 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 257 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 258 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 259 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_xyzw v[0:1], off, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM]
 260 ; GFX11-PACKED-NEXT:    s_nop 0
 261 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 262 ; GFX11-PACKED-NEXT:    s_endpgm
 263 ;
 264 ; GFX12-PACKED-LABEL: tbuffer_store_d16_xyzw:
 265 ; GFX12-PACKED:       ; %bb.0: ; %main_body
 266 ; GFX12-PACKED-NEXT:    s_clause 0x1
 267 ; GFX12-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x34
 268 ; GFX12-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x24
 269 ; GFX12-PACKED-NEXT:    s_wait_kmcnt 0x0
 270 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 271 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 272 ; GFX12-PACKED-NEXT:    tbuffer_store_d16_format_xyzw v[0:1], off, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM]
 273 ; GFX12-PACKED-NEXT:    s_nop 0
 274 ; GFX12-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 275 ; GFX12-PACKED-NEXT:    s_endpgm
 276 main_body:
 277   call void @llvm.amdgcn.raw.tbuffer.store.v4f16(<4 x half> %data, <4 x i32> %rsrc, i32 0, i32 0, i32 33, i32 0)
 278   ret void
 279 }
 280
 281 declare void @llvm.amdgcn.raw.tbuffer.store.f16(half, <4 x i32>, i32, i32, i32, i32)
 282 declare void @llvm.amdgcn.raw.tbuffer.store.v2f16(<2 x half>, <4 x i32>, i32, i32, i32, i32)
 283 declare void @llvm.amdgcn.raw.tbuffer.store.v3f16(<3 x half>, <4 x i32>, i32, i32, i32, i32)
 284 declare void @llvm.amdgcn.raw.tbuffer.store.v4f16(<4 x half>, <4 x i32>, i32, i32, i32, i32)