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

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefixes=PREGFX10-UNPACKED %s
   3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx810 -verify-machineinstrs < %s | FileCheck -check-prefixes=PREGFX10-PACKED %s
   4 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=PREGFX10-PACKED %s
   5 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-PACKED %s
   6 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-PACKED %s
   7 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX12-PACKED,GFX12-PACKED-SDAG %s
   8 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX12-PACKED,GFX12-PACKED-SDAG %s
   9 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX12-PACKED,GFX12-PACKED-GISEL %s
  10
  11 define amdgpu_kernel void @tbuffer_store_d16_x(<4 x i32> %rsrc, half %data, i32 %vindex) {
  12 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_x:
  13 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
  14 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
  15 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
  16 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
  17 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
  18 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v1, s5
  19 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_x v0, v1, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
  20 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
  21 ;
  22 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_x:
  23 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
  24 ; PREGFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
  25 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
  26 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  27 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
  28 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
  29 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_x v0, v1, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
  30 ; PREGFX10-PACKED-NEXT:    s_endpgm
  31 ;
  32 ; GFX10-PACKED-LABEL: tbuffer_store_d16_x:
  33 ; GFX10-PACKED:       ; %bb.0: ; %main_body
  34 ; GFX10-PACKED-NEXT:    s_clause 0x1
  35 ; GFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
  36 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
  37 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  38 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
  39 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
  40 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_x v0, v1, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen
  41 ; GFX10-PACKED-NEXT:    s_endpgm
  42 ;
  43 ; GFX11-PACKED-LABEL: tbuffer_store_d16_x:
  44 ; GFX11-PACKED:       ; %bb.0: ; %main_body
  45 ; GFX11-PACKED-NEXT:    s_clause 0x1
  46 ; GFX11-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x10
  47 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
  48 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  49 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
  50 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v1, s5
  51 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_x v0, v1, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen
  52 ; GFX11-PACKED-NEXT:    s_nop 0
  53 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  54 ; GFX11-PACKED-NEXT:    s_endpgm
  55 ;
  56 ; GFX12-PACKED-LABEL: tbuffer_store_d16_x:
  57 ; GFX12-PACKED:       ; %bb.0: ; %main_body
  58 ; GFX12-PACKED-NEXT:    s_clause 0x1
  59 ; GFX12-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x10
  60 ; GFX12-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
  61 ; GFX12-PACKED-NEXT:    s_wait_kmcnt 0x0
  62 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, s4
  63 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v1, s5
  64 ; GFX12-PACKED-NEXT:    tbuffer_store_d16_format_x v0, v1, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM] idxen
  65 ; GFX12-PACKED-NEXT:    s_nop 0
  66 ; GFX12-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  67 ; GFX12-PACKED-NEXT:    s_endpgm
  68 main_body:
  69   call void @llvm.amdgcn.struct.tbuffer.store.f16(half %data, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0)
  70   ret void
  71 }
  72
  73 define amdgpu_kernel void @tbuffer_store_d16_xy(<4 x i32> %rsrc, <2 x half> %data, i32 %vindex) {
  74 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xy:
  75 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
  76 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
  77 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
  78 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
  79 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s6, s4, 16
  80 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s4, s4, 0xffff
  81 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
  82 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v1, s6
  83 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v2, s5
  84 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_xy v[0:1], v2, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
  85 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
  86 ;
  87 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xy:
  88 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
  89 ; PREGFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
  90 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
  91 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
  92 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
  93 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
  94 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_xy v0, v1, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
  95 ; PREGFX10-PACKED-NEXT:    s_endpgm
  96 ;
  97 ; GFX10-PACKED-LABEL: tbuffer_store_d16_xy:
  98 ; GFX10-PACKED:       ; %bb.0: ; %main_body
  99 ; GFX10-PACKED-NEXT:    s_clause 0x1
 100 ; GFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
 101 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
 102 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 103 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 104 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 105 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_xy v0, v1, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen
 106 ; GFX10-PACKED-NEXT:    s_endpgm
 107 ;
 108 ; GFX11-PACKED-LABEL: tbuffer_store_d16_xy:
 109 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 110 ; GFX11-PACKED-NEXT:    s_clause 0x1
 111 ; GFX11-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x10
 112 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
 113 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 114 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 115 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 116 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_xy v0, v1, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen
 117 ; GFX11-PACKED-NEXT:    s_nop 0
 118 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 119 ; GFX11-PACKED-NEXT:    s_endpgm
 120 ;
 121 ; GFX12-PACKED-LABEL: tbuffer_store_d16_xy:
 122 ; GFX12-PACKED:       ; %bb.0: ; %main_body
 123 ; GFX12-PACKED-NEXT:    s_clause 0x1
 124 ; GFX12-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x10
 125 ; GFX12-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
 126 ; GFX12-PACKED-NEXT:    s_wait_kmcnt 0x0
 127 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 128 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 129 ; GFX12-PACKED-NEXT:    tbuffer_store_d16_format_xy v0, v1, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM] idxen
 130 ; GFX12-PACKED-NEXT:    s_nop 0
 131 ; GFX12-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 132 ; GFX12-PACKED-NEXT:    s_endpgm
 133 main_body:
 134   call void @llvm.amdgcn.struct.tbuffer.store.v2f16(<2 x half> %data, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0)
 135   ret void
 136 }
 137
 138 define amdgpu_kernel void @tbuffer_store_d16_xyz(<4 x i32> %rsrc, <4 x half> %data, i32 %vindex) {
 139 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xyz:
 140 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
 141 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
 142 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
 143 ; PREGFX10-UNPACKED-NEXT:    s_load_dword s6, s[6:7], 0x18
 144 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
 145 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 146 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s7, s4, 16
 147 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s4, s4, 0xffff
 148 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
 149 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v1, s7
 150 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v2, s5
 151 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v3, s6
 152 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_xyz v[0:2], v3, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
 153 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
 154 ;
 155 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xyz:
 156 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
 157 ; PREGFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
 158 ; PREGFX10-PACKED-NEXT:    s_load_dword s8, s[6:7], 0x18
 159 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
 160 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 161 ; PREGFX10-PACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 162 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 163 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 164 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v2, s8
 165 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyz v[0:1], v2, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
 166 ; PREGFX10-PACKED-NEXT:    s_endpgm
 167 ;
 168 ; GFX10-PACKED-LABEL: tbuffer_store_d16_xyz:
 169 ; GFX10-PACKED:       ; %bb.0: ; %main_body
 170 ; GFX10-PACKED-NEXT:    s_clause 0x2
 171 ; GFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
 172 ; GFX10-PACKED-NEXT:    s_load_dword s8, s[6:7], 0x18
 173 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
 174 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 175 ; GFX10-PACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 176 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 177 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 178 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v2, s8
 179 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyz v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen
 180 ; GFX10-PACKED-NEXT:    s_endpgm
 181 ;
 182 ; GFX11-PACKED-LABEL: tbuffer_store_d16_xyz:
 183 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 184 ; GFX11-PACKED-NEXT:    s_clause 0x2
 185 ; GFX11-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x10
 186 ; GFX11-PACKED-NEXT:    s_load_b32 s6, s[2:3], 0x18
 187 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
 188 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 189 ; GFX11-PACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 190 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 191 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 192 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v2, s6
 193 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_xyz v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen
 194 ; GFX11-PACKED-NEXT:    s_nop 0
 195 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 196 ; GFX11-PACKED-NEXT:    s_endpgm
 197 ;
 198 ; GFX12-PACKED-SDAG-LABEL: tbuffer_store_d16_xyz:
 199 ; GFX12-PACKED-SDAG:       ; %bb.0: ; %main_body
 200 ; GFX12-PACKED-SDAG-NEXT:    s_clause 0x1
 201 ; GFX12-PACKED-SDAG-NEXT:    s_load_b96 s[4:6], s[2:3], 0x10
 202 ; GFX12-PACKED-SDAG-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
 203 ; GFX12-PACKED-SDAG-NEXT:    s_wait_kmcnt 0x0
 204 ; GFX12-PACKED-SDAG-NEXT:    s_and_b32 s5, s5, 0xffff
 205 ; GFX12-PACKED-SDAG-NEXT:    v_mov_b32_e32 v0, s4
 206 ; GFX12-PACKED-SDAG-NEXT:    v_mov_b32_e32 v1, s5
 207 ; GFX12-PACKED-SDAG-NEXT:    v_mov_b32_e32 v2, s6
 208 ; GFX12-PACKED-SDAG-NEXT:    tbuffer_store_d16_format_xyz v[0:1], v2, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM] idxen
 209 ; GFX12-PACKED-SDAG-NEXT:    s_nop 0
 210 ; GFX12-PACKED-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 211 ; GFX12-PACKED-SDAG-NEXT:    s_endpgm
 212 ;
 213 ; GFX12-PACKED-GISEL-LABEL: tbuffer_store_d16_xyz:
 214 ; GFX12-PACKED-GISEL:       ; %bb.0: ; %main_body
 215 ; GFX12-PACKED-GISEL-NEXT:    s_clause 0x1
 216 ; GFX12-PACKED-GISEL-NEXT:    s_load_b96 s[4:6], s[2:3], 0x10
 217 ; GFX12-PACKED-GISEL-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
 218 ; GFX12-PACKED-GISEL-NEXT:    s_wait_kmcnt 0x0
 219 ; GFX12-PACKED-GISEL-NEXT:    s_pack_lh_b32_b16 s4, s4, s4
 220 ; GFX12-PACKED-GISEL-NEXT:    v_mov_b32_e32 v2, s6
 221 ; GFX12-PACKED-GISEL-NEXT:    v_mov_b32_e32 v0, s4
 222 ; GFX12-PACKED-GISEL-NEXT:    v_mov_b32_e32 v1, s5
 223 ; GFX12-PACKED-GISEL-NEXT:    tbuffer_store_d16_format_xyzw v[0:1], v2, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM] idxen
 224 ; GFX12-PACKED-GISEL-NEXT:    s_nop 0
 225 ; GFX12-PACKED-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 226 ; GFX12-PACKED-GISEL-NEXT:    s_endpgm
 227 main_body:
 228   %data_subvec = shufflevector <4 x half> %data, <4 x half> undef, <3 x i32> <i32 0, i32 1, i32 2>
 229   call void @llvm.amdgcn.struct.tbuffer.store.v3f16(<3 x half> %data_subvec, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0)
 230   ret void
 231 }
 232
 233 define amdgpu_kernel void @tbuffer_store_d16_xyzw(<4 x i32> %rsrc, <4 x half> %data, i32 %vindex) {
 234 ; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xyzw:
 235 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
 236 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
 237 ; PREGFX10-UNPACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
 238 ; PREGFX10-UNPACKED-NEXT:    s_load_dword s6, s[6:7], 0x18
 239 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt lgkmcnt(0)
 240 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s7, s5, 16
 241 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s5, s5, 0xffff
 242 ; PREGFX10-UNPACKED-NEXT:    s_lshr_b32 s8, s4, 16
 243 ; PREGFX10-UNPACKED-NEXT:    s_and_b32 s4, s4, 0xffff
 244 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, s4
 245 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v1, s8
 246 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v2, s5
 247 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v3, s7
 248 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v4, s6
 249 ; PREGFX10-UNPACKED-NEXT:    tbuffer_store_format_d16_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
 250 ; PREGFX10-UNPACKED-NEXT:    s_endpgm
 251 ;
 252 ; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xyzw:
 253 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
 254 ; PREGFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
 255 ; PREGFX10-PACKED-NEXT:    s_load_dword s8, s[6:7], 0x18
 256 ; PREGFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
 257 ; PREGFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 258 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 259 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 260 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v2, s8
 261 ; PREGFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyzw v[0:1], v2, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen
 262 ; PREGFX10-PACKED-NEXT:    s_endpgm
 263 ;
 264 ; GFX10-PACKED-LABEL: tbuffer_store_d16_xyzw:
 265 ; GFX10-PACKED:       ; %bb.0: ; %main_body
 266 ; GFX10-PACKED-NEXT:    s_clause 0x2
 267 ; GFX10-PACKED-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x10
 268 ; GFX10-PACKED-NEXT:    s_load_dword s8, s[6:7], 0x18
 269 ; GFX10-PACKED-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
 270 ; GFX10-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 271 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 272 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 273 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v2, s8
 274 ; GFX10-PACKED-NEXT:    tbuffer_store_format_d16_xyzw v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen
 275 ; GFX10-PACKED-NEXT:    s_endpgm
 276 ;
 277 ; GFX11-PACKED-LABEL: tbuffer_store_d16_xyzw:
 278 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 279 ; GFX11-PACKED-NEXT:    s_clause 0x2
 280 ; GFX11-PACKED-NEXT:    s_load_b64 s[4:5], s[2:3], 0x10
 281 ; GFX11-PACKED-NEXT:    s_load_b32 s6, s[2:3], 0x18
 282 ; GFX11-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
 283 ; GFX11-PACKED-NEXT:    s_waitcnt lgkmcnt(0)
 284 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 285 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 286 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v2, s6
 287 ; GFX11-PACKED-NEXT:    tbuffer_store_d16_format_xyzw v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen
 288 ; GFX11-PACKED-NEXT:    s_nop 0
 289 ; GFX11-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 290 ; GFX11-PACKED-NEXT:    s_endpgm
 291 ;
 292 ; GFX12-PACKED-LABEL: tbuffer_store_d16_xyzw:
 293 ; GFX12-PACKED:       ; %bb.0: ; %main_body
 294 ; GFX12-PACKED-NEXT:    s_clause 0x1
 295 ; GFX12-PACKED-NEXT:    s_load_b96 s[4:6], s[2:3], 0x10
 296 ; GFX12-PACKED-NEXT:    s_load_b128 s[0:3], s[2:3], 0x0
 297 ; GFX12-PACKED-NEXT:    s_wait_kmcnt 0x0
 298 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, s4
 299 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v1, s5
 300 ; GFX12-PACKED-NEXT:    v_mov_b32_e32 v2, s6
 301 ; GFX12-PACKED-NEXT:    tbuffer_store_d16_format_xyzw v[0:1], v2, s[0:3], null format:[BUF_FMT_10_10_10_2_SNORM] idxen
 302 ; GFX12-PACKED-NEXT:    s_nop 0
 303 ; GFX12-PACKED-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 304 ; GFX12-PACKED-NEXT:    s_endpgm
 305 main_body:
 306   call void @llvm.amdgcn.struct.tbuffer.store.v4f16(<4 x half> %data, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0)
 307   ret void
 308 }
 309
 310 declare void @llvm.amdgcn.struct.tbuffer.store.f16(half, <4 x i32>, i32, i32, i32, i32, i32)
 311 declare void @llvm.amdgcn.struct.tbuffer.store.v2f16(<2 x half>, <4 x i32>, i32, i32, i32, i32, i32)
 312 declare void @llvm.amdgcn.struct.tbuffer.store.v3f16(<3 x half>, <4 x i32>, i32, i32, i32, i32, i32)
 313 declare void @llvm.amdgcn.struct.tbuffer.store.v4f16(<4 x half>, <4 x i32>, i32, i32, i32, i32, i32)