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

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs -show-mc-encoding | FileCheck -enable-var-scope -check-prefixes=PREGFX10-UNPACKED %s
   3 ; RUN: llc < %s -march=amdgcn -mcpu=gfx810 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
   4 ; RUN: llc < %s -march=amdgcn -mcpu=gfx900 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
   5 ; RUN: llc < %s -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX10-PACKED %s
   6 ; RUN: llc < %s -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
   7
   8 define amdgpu_ps half @tbuffer_load_d16_x(<4 x i32> inreg %rsrc) {
   9 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_x:
  10 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
  11 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, 0 ; encoding: [0x80,0x02,0x00,0x7e]
  12 ; PREGFX10-UNPACKED-NEXT:    tbuffer_load_format_d16_x v0, v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen ; encoding: [0x00,0x20,0xb4,0xe8,0x00,0x00,0x00,0x80]
  13 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]
  14 ; PREGFX10-UNPACKED-NEXT:    ; return to shader part epilog
  15 ;
  16 ; PREGFX10-PACKED-LABEL: tbuffer_load_d16_x:
  17 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
  18 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  19 ; PREGFX10-PACKED-NEXT:    tbuffer_load_format_d16_x v0, v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen
  20 ; PREGFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
  21 ; PREGFX10-PACKED-NEXT:    ; return to shader part epilog
  22 ;
  23 ; GFX10-PACKED-LABEL: tbuffer_load_d16_x:
  24 ; GFX10-PACKED:       ; %bb.0: ; %main_body
  25 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  26 ; GFX10-PACKED-NEXT:    tbuffer_load_format_d16_x v0, v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
  27 ; GFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
  28 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
  29 ;
  30 ; GFX11-PACKED-LABEL: tbuffer_load_d16_x:
  31 ; GFX11-PACKED:       ; %bb.0: ; %main_body
  32 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  33 ; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_x v0, v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
  34 ; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
  35 ; GFX11-PACKED-NEXT:    ; return to shader part epilog
  36 main_body:
  37   %data = call half @llvm.amdgcn.struct.tbuffer.load.f16(<4 x i32> %rsrc, i32 0, i32 0, i32 0, i32 22, i32 0)
  38   ret half %data
  39 }
  40
  41 define amdgpu_ps half @tbuffer_load_d16_xy(<4 x i32> inreg %rsrc) {
  42 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_xy:
  43 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
  44 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, 0 ; encoding: [0x80,0x02,0x00,0x7e]
  45 ; PREGFX10-UNPACKED-NEXT:    tbuffer_load_format_d16_xy v[0:1], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen ; encoding: [0x00,0xa0,0xb4,0xe8,0x00,0x00,0x00,0x80]
  46 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]
  47 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, v1 ; encoding: [0x01,0x03,0x00,0x7e]
  48 ; PREGFX10-UNPACKED-NEXT:    ; return to shader part epilog
  49 ;
  50 ; PREGFX10-PACKED-LABEL: tbuffer_load_d16_xy:
  51 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
  52 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  53 ; PREGFX10-PACKED-NEXT:    tbuffer_load_format_d16_xy v0, v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen
  54 ; PREGFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
  55 ; PREGFX10-PACKED-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
  56 ; PREGFX10-PACKED-NEXT:    ; return to shader part epilog
  57 ;
  58 ; GFX10-PACKED-LABEL: tbuffer_load_d16_xy:
  59 ; GFX10-PACKED:       ; %bb.0: ; %main_body
  60 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  61 ; GFX10-PACKED-NEXT:    tbuffer_load_format_d16_xy v0, v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
  62 ; GFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
  63 ; GFX10-PACKED-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
  64 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
  65 ;
  66 ; GFX11-PACKED-LABEL: tbuffer_load_d16_xy:
  67 ; GFX11-PACKED:       ; %bb.0: ; %main_body
  68 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  69 ; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_xy v0, v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
  70 ; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
  71 ; GFX11-PACKED-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
  72 ; GFX11-PACKED-NEXT:    ; return to shader part epilog
  73 main_body:
  74   %data = call <2 x half> @llvm.amdgcn.struct.tbuffer.load.v2f16(<4 x i32> %rsrc, i32 0, i32 0, i32 0, i32 22, i32 0)
  75   %elt = extractelement <2 x half> %data, i32 1
  76   ret half %elt
  77 }
  78
  79 define amdgpu_ps half @tbuffer_load_d16_xyz(<4 x i32> inreg %rsrc) {
  80 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_xyz:
  81 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
  82 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, 0 ; encoding: [0x80,0x02,0x00,0x7e]
  83 ; PREGFX10-UNPACKED-NEXT:    tbuffer_load_format_d16_xyz v[0:2], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen ; encoding: [0x00,0x20,0xb5,0xe8,0x00,0x00,0x00,0x80]
  84 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]
  85 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, v2 ; encoding: [0x02,0x03,0x00,0x7e]
  86 ; PREGFX10-UNPACKED-NEXT:    ; return to shader part epilog
  87 ;
  88 ; PREGFX10-PACKED-LABEL: tbuffer_load_d16_xyz:
  89 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
  90 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  91 ; PREGFX10-PACKED-NEXT:    tbuffer_load_format_d16_xyz v[0:1], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen
  92 ; PREGFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
  93 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, v1
  94 ; PREGFX10-PACKED-NEXT:    ; return to shader part epilog
  95 ;
  96 ; GFX10-PACKED-LABEL: tbuffer_load_d16_xyz:
  97 ; GFX10-PACKED:       ; %bb.0: ; %main_body
  98 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
  99 ; GFX10-PACKED-NEXT:    tbuffer_load_format_d16_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
 100 ; GFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
 101 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, v1
 102 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
 103 ;
 104 ; GFX11-PACKED-LABEL: tbuffer_load_d16_xyz:
 105 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 106 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, 0
 107 ; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
 108 ; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
 109 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, v1
 110 ; GFX11-PACKED-NEXT:    ; return to shader part epilog
 111 main_body:
 112   %data = call <3 x half> @llvm.amdgcn.struct.tbuffer.load.v3f16(<4 x i32> %rsrc, i32 0, i32 0, i32 0, i32 22, i32 0)
 113   %elt = extractelement <3 x half> %data, i32 2
 114   ret half %elt
 115 }
 116
 117 define amdgpu_ps half @tbuffer_load_d16_xyzw(<4 x i32> inreg %rsrc) {
 118 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_xyzw:
 119 ; PREGFX10-UNPACKED:       ; %bb.0: ; %main_body
 120 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, 0 ; encoding: [0x80,0x02,0x00,0x7e]
 121 ; PREGFX10-UNPACKED-NEXT:    tbuffer_load_format_d16_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen ; encoding: [0x00,0xa0,0xb5,0xe8,0x00,0x00,0x00,0x80]
 122 ; PREGFX10-UNPACKED-NEXT:    s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]
 123 ; PREGFX10-UNPACKED-NEXT:    v_mov_b32_e32 v0, v3 ; encoding: [0x03,0x03,0x00,0x7e]
 124 ; PREGFX10-UNPACKED-NEXT:    ; return to shader part epilog
 125 ;
 126 ; PREGFX10-PACKED-LABEL: tbuffer_load_d16_xyzw:
 127 ; PREGFX10-PACKED:       ; %bb.0: ; %main_body
 128 ; PREGFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
 129 ; PREGFX10-PACKED-NEXT:    tbuffer_load_format_d16_xyzw v[0:1], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] idxen
 130 ; PREGFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
 131 ; PREGFX10-PACKED-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
 132 ; PREGFX10-PACKED-NEXT:    ; return to shader part epilog
 133 ;
 134 ; GFX10-PACKED-LABEL: tbuffer_load_d16_xyzw:
 135 ; GFX10-PACKED:       ; %bb.0: ; %main_body
 136 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, 0
 137 ; GFX10-PACKED-NEXT:    tbuffer_load_format_d16_xyzw v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
 138 ; GFX10-PACKED-NEXT:    s_waitcnt vmcnt(0)
 139 ; GFX10-PACKED-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
 140 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
 141 ;
 142 ; GFX11-PACKED-LABEL: tbuffer_load_d16_xyzw:
 143 ; GFX11-PACKED:       ; %bb.0: ; %main_body
 144 ; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, 0
 145 ; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_xyzw v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
 146 ; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
 147 ; GFX11-PACKED-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
 148 ; GFX11-PACKED-NEXT:    ; return to shader part epilog
 149 main_body:
 150   %data = call <4 x half> @llvm.amdgcn.struct.tbuffer.load.v4f16(<4 x i32> %rsrc, i32 0, i32 0, i32 0, i32 22, i32 0)
 151   %elt = extractelement <4 x half> %data, i32 3
 152   ret half %elt
 153 }
 154
 155 declare half @llvm.amdgcn.struct.tbuffer.load.f16(<4 x i32>, i32, i32, i32, i32, i32)
 156 declare <2 x half> @llvm.amdgcn.struct.tbuffer.load.v2f16(<4 x i32>, i32, i32, i32, i32, i32)
 157 declare <3 x half> @llvm.amdgcn.struct.tbuffer.load.v3f16(<4 x i32>, i32, i32, i32, i32, i32)
 158 declare <4 x half> @llvm.amdgcn.struct.tbuffer.load.v4f16(<4 x i32>, i32, i32, i32, i32, i32)