llvm/test/CodeGen/AMDGPU/fptrunc.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
   2 ; RUN: llc -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefixes=SI %s
   3 ; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-SDAG,VI-SAFE-SDAG %s
   4 ; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-GISEL,VI-SAFE-GISEL %s
   5 ; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-SDAG,VI-UNSAFE-SDAG %s
   6 ; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-GISEL,VI-UNSAFE-GISEL %s
   7 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-SAFE-SDAG %s
   8 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-GISEL,GFX10-SAFE-GISEL %s
   9 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-UNSAFE-SDAG %s
  10 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-GISEL,GFX10-UNSAFE-GISEL %s
  11 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-SAFE-SDAG %s
  12 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-SAFE-GISEL %s
  13 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-UNSAFE-SDAG %s
  14 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-UNSAFE-GISEL %s
  15
  16 define amdgpu_kernel void @fptrunc_f64_to_f32(ptr addrspace(1) %out, double %in) {
  17 ; SI-LABEL: fptrunc_f64_to_f32:
  18 ; SI:       ; %bb.0:
  19 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
  20 ; SI-NEXT:    s_mov_b32 s7, 0xf000
  21 ; SI-NEXT:    s_mov_b32 s6, -1
  22 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
  23 ; SI-NEXT:    s_mov_b32 s4, s0
  24 ; SI-NEXT:    s_mov_b32 s5, s1
  25 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  26 ; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
  27 ; SI-NEXT:    s_endpgm
  28 ;
  29 ; VI-SDAG-LABEL: fptrunc_f64_to_f32:
  30 ; VI-SDAG:       ; %bb.0:
  31 ; VI-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
  32 ; VI-SDAG-NEXT:    s_mov_b32 s7, 0xf000
  33 ; VI-SDAG-NEXT:    s_mov_b32 s6, -1
  34 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  35 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  36 ; VI-SDAG-NEXT:    s_mov_b32 s4, s0
  37 ; VI-SDAG-NEXT:    s_mov_b32 s5, s1
  38 ; VI-SDAG-NEXT:    buffer_store_dword v0, off, s[4:7], 0
  39 ; VI-SDAG-NEXT:    s_endpgm
  40 ;
  41 ; VI-GISEL-LABEL: fptrunc_f64_to_f32:
  42 ; VI-GISEL:       ; %bb.0:
  43 ; VI-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
  44 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
  45 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  46 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
  47 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
  48 ; VI-GISEL-NEXT:    buffer_store_dword v0, off, s[0:3], 0
  49 ; VI-GISEL-NEXT:    s_endpgm
  50 ;
  51 ; GFX10-SDAG-LABEL: fptrunc_f64_to_f32:
  52 ; GFX10-SDAG:       ; %bb.0:
  53 ; GFX10-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
  54 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  55 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  56 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
  57 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
  58 ; GFX10-SDAG-NEXT:    buffer_store_dword v0, off, s[0:3], 0
  59 ; GFX10-SDAG-NEXT:    s_endpgm
  60 ;
  61 ; GFX10-GISEL-LABEL: fptrunc_f64_to_f32:
  62 ; GFX10-GISEL:       ; %bb.0:
  63 ; GFX10-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
  64 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
  65 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  66 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
  67 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
  68 ; GFX10-GISEL-NEXT:    buffer_store_dword v0, off, s[0:3], 0
  69 ; GFX10-GISEL-NEXT:    s_endpgm
  70 ;
  71 ; GFX11-SDAG-LABEL: fptrunc_f64_to_f32:
  72 ; GFX11-SDAG:       ; %bb.0:
  73 ; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
  74 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  75 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  76 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
  77 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
  78 ; GFX11-SDAG-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
  79 ; GFX11-SDAG-NEXT:    s_endpgm
  80 ;
  81 ; GFX11-GISEL-LABEL: fptrunc_f64_to_f32:
  82 ; GFX11-GISEL:       ; %bb.0:
  83 ; GFX11-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
  84 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
  85 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  86 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
  87 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
  88 ; GFX11-GISEL-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
  89 ; GFX11-GISEL-NEXT:    s_endpgm
  90   %result = fptrunc double %in to float
  91   store float %result, ptr addrspace(1) %out
  92   ret void
  93 }
  94
  95 define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) {
  96 ; SI-LABEL: fptrunc_f64_to_f16:
  97 ; SI:       ; %bb.0:
  98 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x9
  99 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 100 ; SI-NEXT:    s_mov_b32 s2, -1
 101 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 102 ; SI-NEXT:    s_mov_b32 s0, s4
 103 ; SI-NEXT:    s_mov_b32 s1, s5
 104 ; SI-NEXT:    s_lshr_b32 s4, s7, 8
 105 ; SI-NEXT:    s_and_b32 s5, s7, 0x1ff
 106 ; SI-NEXT:    s_and_b32 s8, s4, 0xffe
 107 ; SI-NEXT:    s_or_b32 s4, s5, s6
 108 ; SI-NEXT:    s_cmp_lg_u32 s4, 0
 109 ; SI-NEXT:    s_cselect_b64 s[4:5], -1, 0
 110 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 111 ; SI-NEXT:    s_bfe_u32 s4, s7, 0xb0014
 112 ; SI-NEXT:    v_readfirstlane_b32 s5, v0
 113 ; SI-NEXT:    s_sub_i32 s6, 0x3f1, s4
 114 ; SI-NEXT:    s_add_i32 s10, s4, 0xfffffc10
 115 ; SI-NEXT:    s_or_b32 s11, s8, s5
 116 ; SI-NEXT:    v_med3_i32 v0, s6, 0, 13
 117 ; SI-NEXT:    s_lshl_b32 s4, s10, 12
 118 ; SI-NEXT:    s_or_b32 s5, s11, 0x1000
 119 ; SI-NEXT:    v_readfirstlane_b32 s6, v0
 120 ; SI-NEXT:    s_or_b32 s4, s11, s4
 121 ; SI-NEXT:    s_lshr_b32 s6, s5, s6
 122 ; SI-NEXT:    v_lshl_b32_e32 v0, s6, v0
 123 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, s5, v0
 124 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
 125 ; SI-NEXT:    v_readfirstlane_b32 s5, v0
 126 ; SI-NEXT:    s_or_b32 s5, s6, s5
 127 ; SI-NEXT:    s_cmp_lt_i32 s10, 1
 128 ; SI-NEXT:    s_cselect_b32 s6, s5, s4
 129 ; SI-NEXT:    s_and_b32 s8, s6, 7
 130 ; SI-NEXT:    s_cmp_gt_i32 s8, 5
 131 ; SI-NEXT:    s_cselect_b64 s[4:5], -1, 0
 132 ; SI-NEXT:    s_cmp_eq_u32 s8, 3
 133 ; SI-NEXT:    s_cselect_b64 s[8:9], -1, 0
 134 ; SI-NEXT:    s_lshr_b32 s6, s6, 2
 135 ; SI-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
 136 ; SI-NEXT:    s_or_b32 s4, s4, s5
 137 ; SI-NEXT:    s_cmp_lg_u32 s4, 0
 138 ; SI-NEXT:    s_addc_u32 s4, s6, 0
 139 ; SI-NEXT:    s_cmp_lt_i32 s10, 31
 140 ; SI-NEXT:    s_cselect_b32 s6, s4, 0x7c00
 141 ; SI-NEXT:    s_cmp_lg_u32 s11, 0
 142 ; SI-NEXT:    s_cselect_b64 s[4:5], -1, 0
 143 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 144 ; SI-NEXT:    s_cmpk_eq_i32 s10, 0x40f
 145 ; SI-NEXT:    v_mov_b32_e32 v1, s6
 146 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 147 ; SI-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 148 ; SI-NEXT:    s_cselect_b64 vcc, -1, 0
 149 ; SI-NEXT:    s_lshr_b32 s4, s7, 16
 150 ; SI-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 151 ; SI-NEXT:    s_and_b32 s4, s4, 0x8000
 152 ; SI-NEXT:    v_or_b32_e32 v0, s4, v0
 153 ; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
 154 ; SI-NEXT:    s_endpgm
 155 ;
 156 ; VI-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 157 ; VI-SAFE-SDAG:       ; %bb.0:
 158 ; VI-SAFE-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x24
 159 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 160 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 161 ; VI-SAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 162 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s0, s4
 163 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s4, s7, 8
 164 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s8, s4, 0xffe
 165 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s4, s7, 0x1ff
 166 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s4, s4, s6
 167 ; VI-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s4, 0
 168 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s1, s5
 169 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[4:5], -1, 0
 170 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 171 ; VI-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
 172 ; VI-SAFE-SDAG-NEXT:    s_bfe_u32 s5, s7, 0xb0014
 173 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s6, s8, s4
 174 ; VI-SAFE-SDAG-NEXT:    s_sub_i32 s8, 0x3f1, s5
 175 ; VI-SAFE-SDAG-NEXT:    v_med3_i32 v0, s8, 0, 13
 176 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s4, s6, 0x1000
 177 ; VI-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s8, v0
 178 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s8, s4, s8
 179 ; VI-SAFE-SDAG-NEXT:    v_lshlrev_b32_e64 v0, v0, s8
 180 ; VI-SAFE-SDAG-NEXT:    v_cmp_ne_u32_e32 vcc, s4, v0
 181 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
 182 ; VI-SAFE-SDAG-NEXT:    s_add_i32 s10, s5, 0xfffffc10
 183 ; VI-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
 184 ; VI-SAFE-SDAG-NEXT:    s_lshl_b32 s5, s10, 12
 185 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s4, s8, s4
 186 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s5, s6, s5
 187 ; VI-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s10, 1
 188 ; VI-SAFE-SDAG-NEXT:    s_cselect_b32 s11, s4, s5
 189 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s8, s11, 7
 190 ; VI-SAFE-SDAG-NEXT:    s_cmp_gt_i32 s8, 5
 191 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[4:5], -1, 0
 192 ; VI-SAFE-SDAG-NEXT:    s_cmp_eq_u32 s8, 3
 193 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[8:9], -1, 0
 194 ; VI-SAFE-SDAG-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
 195 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s8, s11, 2
 196 ; VI-SAFE-SDAG-NEXT:    s_cmp_lg_u64 s[4:5], 0
 197 ; VI-SAFE-SDAG-NEXT:    s_addc_u32 s4, s8, 0
 198 ; VI-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s10, 31
 199 ; VI-SAFE-SDAG-NEXT:    s_cselect_b32 s8, s4, 0x7c00
 200 ; VI-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s6, 0
 201 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[4:5], -1, 0
 202 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 203 ; VI-SAFE-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 204 ; VI-SAFE-SDAG-NEXT:    s_cmpk_eq_i32 s10, 0x40f
 205 ; VI-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 206 ; VI-SAFE-SDAG-NEXT:    v_mov_b32_e32 v1, s8
 207 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 vcc, -1, 0
 208 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s4, s7, 16
 209 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 210 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s4, s4, 0x8000
 211 ; VI-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, s4, v0
 212 ; VI-SAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 213 ; VI-SAFE-SDAG-NEXT:    s_endpgm
 214 ;
 215 ; VI-SAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 216 ; VI-SAFE-GISEL:       ; %bb.0:
 217 ; VI-SAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 218 ; VI-SAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 219 ; VI-SAFE-GISEL-NEXT:    s_bfe_u32 s4, s3, 0xb0014
 220 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s5, s3, 8
 221 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s6, s3, 0x1ff
 222 ; VI-SAFE-GISEL-NEXT:    s_addk_i32 s4, 0xfc10
 223 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s5, s5, 0xffe
 224 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s6, s2
 225 ; VI-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 226 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 227 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s5, s2
 228 ; VI-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 229 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s5, 1, 0
 230 ; VI-SAFE-GISEL-NEXT:    s_sub_i32 s7, 1, s4
 231 ; VI-SAFE-GISEL-NEXT:    s_lshl_b32 s6, s4, 12
 232 ; VI-SAFE-GISEL-NEXT:    s_max_i32 s7, s7, 0
 233 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s6, s2, s6
 234 ; VI-SAFE-GISEL-NEXT:    s_min_i32 s7, s7, 13
 235 ; VI-SAFE-GISEL-NEXT:    s_bitset1_b32 s2, 12
 236 ; VI-SAFE-GISEL-NEXT:    s_lshl_b32 s5, s5, 9
 237 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s8, s2, s7
 238 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s5, s5, 0x7c00
 239 ; VI-SAFE-GISEL-NEXT:    s_lshl_b32 s7, s8, s7
 240 ; VI-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s7, s2
 241 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 242 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s8, s2
 243 ; VI-SAFE-GISEL-NEXT:    s_cmp_lt_i32 s4, 1
 244 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s2, s6
 245 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s6, s2, 7
 246 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s2, s2, 2
 247 ; VI-SAFE-GISEL-NEXT:    s_cmp_eq_u32 s6, 3
 248 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s7, 1, 0
 249 ; VI-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s6, 5
 250 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 251 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s6, s7, s6
 252 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s6, s6, 1
 253 ; VI-SAFE-GISEL-NEXT:    s_add_i32 s2, s2, s6
 254 ; VI-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s4, 30
 255 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 0x7c00, s2
 256 ; VI-SAFE-GISEL-NEXT:    s_cmpk_eq_i32 s4, 0x40f
 257 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s5, s2
 258 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s3, s3, 16
 259 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s3, s3, 0x8000
 260 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s3, s2
 261 ; VI-SAFE-GISEL-NEXT:    v_mov_b32_e32 v0, s2
 262 ; VI-SAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 263 ; VI-SAFE-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 264 ; VI-SAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 265 ; VI-SAFE-GISEL-NEXT:    s_endpgm
 266 ;
 267 ; VI-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 268 ; VI-UNSAFE-SDAG:       ; %bb.0:
 269 ; VI-UNSAFE-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 270 ; VI-UNSAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 271 ; VI-UNSAFE-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 272 ; VI-UNSAFE-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 273 ; VI-UNSAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 274 ; VI-UNSAFE-SDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 275 ; VI-UNSAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 276 ; VI-UNSAFE-SDAG-NEXT:    s_endpgm
 277 ;
 278 ; VI-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 279 ; VI-UNSAFE-GISEL:       ; %bb.0:
 280 ; VI-UNSAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 281 ; VI-UNSAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 282 ; VI-UNSAFE-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 283 ; VI-UNSAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 284 ; VI-UNSAFE-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 285 ; VI-UNSAFE-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
 286 ; VI-UNSAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 287 ; VI-UNSAFE-GISEL-NEXT:    s_endpgm
 288 ;
 289 ; GFX10-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 290 ; GFX10-SAFE-SDAG:       ; %bb.0:
 291 ; GFX10-SAFE-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 292 ; GFX10-SAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 293 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s4, s3, 0x1ff
 294 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s3, 8
 295 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s2, s4, s2
 296 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s4, s5, 0xffe
 297 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s2, 0
 298 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s2, -1, 0
 299 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s2
 300 ; GFX10-SAFE-SDAG-NEXT:    s_bfe_u32 s2, s3, 0xb0014
 301 ; GFX10-SAFE-SDAG-NEXT:    s_sub_i32 s5, 0x3f1, s2
 302 ; GFX10-SAFE-SDAG-NEXT:    s_addk_i32 s2, 0xfc10
 303 ; GFX10-SAFE-SDAG-NEXT:    v_med3_i32 v1, s5, 0, 13
 304 ; GFX10-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 305 ; GFX10-SAFE-SDAG-NEXT:    s_lshl_b32 s7, s2, 12
 306 ; GFX10-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s6, v1
 307 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s4, s4, s5
 308 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s5, s4, 0x1000
 309 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s7, s4, s7
 310 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s6, s5, s6
 311 ; GFX10-SAFE-SDAG-NEXT:    v_lshlrev_b32_e64 v0, v1, s6
 312 ; GFX10-SAFE-SDAG-NEXT:    v_cmp_ne_u32_e32 vcc_lo, s5, v0
 313 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc_lo
 314 ; GFX10-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 315 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s5, s6, s5
 316 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 1
 317 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, s7
 318 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s6, s5, 7
 319 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_gt_i32 s6, 5
 320 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s7, -1, 0
 321 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_eq_u32 s6, 3
 322 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s6, -1, 0
 323 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s5, 2
 324 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s6, s6, s7
 325 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s6, 0
 326 ; GFX10-SAFE-SDAG-NEXT:    s_addc_u32 s5, s5, 0
 327 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 31
 328 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, 0x7c00
 329 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s4, 0
 330 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s4, -1, 0
 331 ; GFX10-SAFE-SDAG-NEXT:    s_cmpk_eq_i32 s2, 0x40f
 332 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s4
 333 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 vcc_lo, -1, 0
 334 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s2, s3, 16
 335 ; GFX10-SAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 336 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s2, s2, 0x8000
 337 ; GFX10-SAFE-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 338 ; GFX10-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 339 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e32 v0, s5, v0, vcc_lo
 340 ; GFX10-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, s2, v0
 341 ; GFX10-SAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 342 ; GFX10-SAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 343 ; GFX10-SAFE-SDAG-NEXT:    s_endpgm
 344 ;
 345 ; GFX10-SAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 346 ; GFX10-SAFE-GISEL:       ; %bb.0:
 347 ; GFX10-SAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 348 ; GFX10-SAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 349 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s6, s3, 0x1ff
 350 ; GFX10-SAFE-GISEL-NEXT:    s_bfe_u32 s4, s3, 0xb0014
 351 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s5, s3, 8
 352 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s6, s2
 353 ; GFX10-SAFE-GISEL-NEXT:    s_addk_i32 s4, 0xfc10
 354 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s5, s5, 0xffe
 355 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 356 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 357 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s5, s2
 358 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 359 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s5, 1, 0
 360 ; GFX10-SAFE-GISEL-NEXT:    s_sub_i32 s6, 1, s4
 361 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s8, s2, 0x1000
 362 ; GFX10-SAFE-GISEL-NEXT:    s_max_i32 s6, s6, 0
 363 ; GFX10-SAFE-GISEL-NEXT:    s_lshl_b32 s7, s4, 12
 364 ; GFX10-SAFE-GISEL-NEXT:    s_min_i32 s6, s6, 13
 365 ; GFX10-SAFE-GISEL-NEXT:    s_lshl_b32 s5, s5, 9
 366 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s9, s8, s6
 367 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s2, s7
 368 ; GFX10-SAFE-GISEL-NEXT:    s_lshl_b32 s6, s9, s6
 369 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s5, s5, 0x7c00
 370 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s6, s8
 371 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 372 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s6, s9, s6
 373 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lt_i32 s4, 1
 374 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s6, s2
 375 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s6, s2, 7
 376 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s2, s2, 2
 377 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_eq_u32 s6, 3
 378 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s7, 1, 0
 379 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s6, 5
 380 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 381 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s6, s7, s6
 382 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s6, s6, 1
 383 ; GFX10-SAFE-GISEL-NEXT:    s_add_i32 s2, s2, s6
 384 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s4, 30
 385 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 0x7c00, s2
 386 ; GFX10-SAFE-GISEL-NEXT:    s_cmpk_eq_i32 s4, 0x40f
 387 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s5, s2
 388 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s3, s3, 16
 389 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s3, s3, 0x8000
 390 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s3, s2
 391 ; GFX10-SAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 392 ; GFX10-SAFE-GISEL-NEXT:    v_mov_b32_e32 v0, s2
 393 ; GFX10-SAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 394 ; GFX10-SAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 395 ; GFX10-SAFE-GISEL-NEXT:    s_endpgm
 396 ;
 397 ; GFX10-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 398 ; GFX10-UNSAFE-SDAG:       ; %bb.0:
 399 ; GFX10-UNSAFE-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 400 ; GFX10-UNSAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 401 ; GFX10-UNSAFE-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 402 ; GFX10-UNSAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 403 ; GFX10-UNSAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 404 ; GFX10-UNSAFE-SDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 405 ; GFX10-UNSAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 406 ; GFX10-UNSAFE-SDAG-NEXT:    s_endpgm
 407 ;
 408 ; GFX10-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 409 ; GFX10-UNSAFE-GISEL:       ; %bb.0:
 410 ; GFX10-UNSAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 411 ; GFX10-UNSAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 412 ; GFX10-UNSAFE-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 413 ; GFX10-UNSAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 414 ; GFX10-UNSAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 415 ; GFX10-UNSAFE-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
 416 ; GFX10-UNSAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 417 ; GFX10-UNSAFE-GISEL-NEXT:    s_endpgm
 418 ;
 419 ; GFX11-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 420 ; GFX11-SAFE-SDAG:       ; %bb.0:
 421 ; GFX11-SAFE-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 422 ; GFX11-SAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 423 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s4, s3, 0x1ff
 424 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s3, 8
 425 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s2, s4, s2
 426 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s4, s5, 0xffe
 427 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s2, 0
 428 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s2, -1, 0
 429 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
 430 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s2
 431 ; GFX11-SAFE-SDAG-NEXT:    s_bfe_u32 s2, s3, 0xb0014
 432 ; GFX11-SAFE-SDAG-NEXT:    s_sub_i32 s5, 0x3f1, s2
 433 ; GFX11-SAFE-SDAG-NEXT:    s_addk_i32 s2, 0xfc10
 434 ; GFX11-SAFE-SDAG-NEXT:    v_med3_i32 v1, s5, 0, 13
 435 ; GFX11-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 436 ; GFX11-SAFE-SDAG-NEXT:    s_lshl_b32 s7, s2, 12
 437 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 438 ; GFX11-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s6, v1
 439 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s4, s4, s5
 440 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 441 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s5, s4, 0x1000
 442 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s7, s4, s7
 443 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s6, s5, s6
 444 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 445 ; GFX11-SAFE-SDAG-NEXT:    v_lshlrev_b32_e64 v0, v1, s6
 446 ; GFX11-SAFE-SDAG-NEXT:    v_cmp_ne_u32_e32 vcc_lo, s5, v0
 447 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc_lo
 448 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 449 ; GFX11-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 450 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s5, s6, s5
 451 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 1
 452 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, s7
 453 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 454 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s6, s5, 7
 455 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_gt_i32 s6, 5
 456 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s7, -1, 0
 457 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_eq_u32 s6, 3
 458 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s6, -1, 0
 459 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s5, 2
 460 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s6, s6, s7
 461 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 462 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s6, 0
 463 ; GFX11-SAFE-SDAG-NEXT:    s_addc_u32 s5, s5, 0
 464 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 31
 465 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, 0x7c00
 466 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s4, 0
 467 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s4, -1, 0
 468 ; GFX11-SAFE-SDAG-NEXT:    s_cmpk_eq_i32 s2, 0x40f
 469 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s4
 470 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 vcc_lo, -1, 0
 471 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s2, s3, 16
 472 ; GFX11-SAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 473 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s2, s2, 0x8000
 474 ; GFX11-SAFE-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 475 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 476 ; GFX11-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 477 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e32 v0, s5, v0, vcc_lo
 478 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 479 ; GFX11-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, s2, v0
 480 ; GFX11-SAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 481 ; GFX11-SAFE-SDAG-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 482 ; GFX11-SAFE-SDAG-NEXT:    s_endpgm
 483 ;
 484 ; GFX11-SAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 485 ; GFX11-SAFE-GISEL:       ; %bb.0:
 486 ; GFX11-SAFE-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 487 ; GFX11-SAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 488 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s6, s3, 0x1ff
 489 ; GFX11-SAFE-GISEL-NEXT:    s_bfe_u32 s4, s3, 0xb0014
 490 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s5, s3, 8
 491 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s6, s2
 492 ; GFX11-SAFE-GISEL-NEXT:    s_addk_i32 s4, 0xfc10
 493 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s5, s5, 0xffe
 494 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 495 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 496 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 497 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s5, s2
 498 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 499 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s5, 1, 0
 500 ; GFX11-SAFE-GISEL-NEXT:    s_sub_i32 s6, 1, s4
 501 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s8, s2, 0x1000
 502 ; GFX11-SAFE-GISEL-NEXT:    s_max_i32 s6, s6, 0
 503 ; GFX11-SAFE-GISEL-NEXT:    s_lshl_b32 s7, s4, 12
 504 ; GFX11-SAFE-GISEL-NEXT:    s_min_i32 s6, s6, 13
 505 ; GFX11-SAFE-GISEL-NEXT:    s_lshl_b32 s5, s5, 9
 506 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s9, s8, s6
 507 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s2, s7
 508 ; GFX11-SAFE-GISEL-NEXT:    s_lshl_b32 s6, s9, s6
 509 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s5, s5, 0x7c00
 510 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s6, s8
 511 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 512 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
 513 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s6, s9, s6
 514 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lt_i32 s4, 1
 515 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s6, s2
 516 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s6, s2, 7
 517 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s2, s2, 2
 518 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_eq_u32 s6, 3
 519 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s7, 1, 0
 520 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s6, 5
 521 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 522 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 523 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s6, s7, s6
 524 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s6, s6, 1
 525 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 526 ; GFX11-SAFE-GISEL-NEXT:    s_add_i32 s2, s2, s6
 527 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s4, 30
 528 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 0x7c00, s2
 529 ; GFX11-SAFE-GISEL-NEXT:    s_cmpk_eq_i32 s4, 0x40f
 530 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s5, s2
 531 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s3, s3, 16
 532 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 533 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s3, s3, 0x8000
 534 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s3, s2
 535 ; GFX11-SAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 536 ; GFX11-SAFE-GISEL-NEXT:    v_mov_b32_e32 v0, s2
 537 ; GFX11-SAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 538 ; GFX11-SAFE-GISEL-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 539 ; GFX11-SAFE-GISEL-NEXT:    s_endpgm
 540 ;
 541 ; GFX11-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 542 ; GFX11-UNSAFE-SDAG:       ; %bb.0:
 543 ; GFX11-UNSAFE-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 544 ; GFX11-UNSAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 545 ; GFX11-UNSAFE-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 546 ; GFX11-UNSAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 547 ; GFX11-UNSAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 548 ; GFX11-UNSAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 549 ; GFX11-UNSAFE-SDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 550 ; GFX11-UNSAFE-SDAG-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 551 ; GFX11-UNSAFE-SDAG-NEXT:    s_endpgm
 552 ;
 553 ; GFX11-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 554 ; GFX11-UNSAFE-GISEL:       ; %bb.0:
 555 ; GFX11-UNSAFE-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 556 ; GFX11-UNSAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 557 ; GFX11-UNSAFE-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 558 ; GFX11-UNSAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 559 ; GFX11-UNSAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 560 ; GFX11-UNSAFE-GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 561 ; GFX11-UNSAFE-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
 562 ; GFX11-UNSAFE-GISEL-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 563 ; GFX11-UNSAFE-GISEL-NEXT:    s_endpgm
 564   %result = fptrunc double %in to half
 565   %result_i16 = bitcast half %result to i16
 566   store i16 %result_i16, ptr addrspace(1) %out
 567   ret void
 568 }
 569
 570 define amdgpu_kernel void @fptrunc_v2f64_to_v2f32(ptr addrspace(1) %out, <2 x double> %in) {
 571 ; SI-LABEL: fptrunc_v2f64_to_v2f32:
 572 ; SI:       ; %bb.0:
 573 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xd
 574 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x9
 575 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 576 ; SI-NEXT:    s_mov_b32 s6, -1
 577 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 578 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 579 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 580 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 581 ; SI-NEXT:    s_endpgm
 582 ;
 583 ; VI-SDAG-LABEL: fptrunc_v2f64_to_v2f32:
 584 ; VI-SDAG:       ; %bb.0:
 585 ; VI-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x34
 586 ; VI-SDAG-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x24
 587 ; VI-SDAG-NEXT:    s_mov_b32 s7, 0xf000
 588 ; VI-SDAG-NEXT:    s_mov_b32 s6, -1
 589 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 590 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 591 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 592 ; VI-SDAG-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 593 ; VI-SDAG-NEXT:    s_endpgm
 594 ;
 595 ; VI-GISEL-LABEL: fptrunc_v2f64_to_v2f32:
 596 ; VI-GISEL:       ; %bb.0:
 597 ; VI-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x34
 598 ; VI-GISEL-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x24
 599 ; VI-GISEL-NEXT:    s_mov_b32 s6, -1
 600 ; VI-GISEL-NEXT:    s_mov_b32 s7, 0xf000
 601 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 602 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 603 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 604 ; VI-GISEL-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 605 ; VI-GISEL-NEXT:    s_endpgm
 606 ;
 607 ; GFX10-SDAG-LABEL: fptrunc_v2f64_to_v2f32:
 608 ; GFX10-SDAG:       ; %bb.0:
 609 ; GFX10-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x34
 610 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 611 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 612 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 613 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 614 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 615 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 616 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 617 ; GFX10-SDAG-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 618 ; GFX10-SDAG-NEXT:    s_endpgm
 619 ;
 620 ; GFX10-GISEL-LABEL: fptrunc_v2f64_to_v2f32:
 621 ; GFX10-GISEL:       ; %bb.0:
 622 ; GFX10-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x34
 623 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 624 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 625 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 626 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 627 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 628 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 629 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 630 ; GFX10-GISEL-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 631 ; GFX10-GISEL-NEXT:    s_endpgm
 632 ;
 633 ; GFX11-SDAG-LABEL: fptrunc_v2f64_to_v2f32:
 634 ; GFX11-SDAG:       ; %bb.0:
 635 ; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x34
 636 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 637 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 638 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 639 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
 640 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 641 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
 642 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 643 ; GFX11-SDAG-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
 644 ; GFX11-SDAG-NEXT:    s_endpgm
 645 ;
 646 ; GFX11-GISEL-LABEL: fptrunc_v2f64_to_v2f32:
 647 ; GFX11-GISEL:       ; %bb.0:
 648 ; GFX11-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x34
 649 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 650 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 651 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 652 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
 653 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
 654 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 655 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 656 ; GFX11-GISEL-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
 657 ; GFX11-GISEL-NEXT:    s_endpgm
 658   %result = fptrunc <2 x double> %in to <2 x float>
 659   store <2 x float> %result, ptr addrspace(1) %out
 660   ret void
 661 }
 662
 663 define amdgpu_kernel void @fptrunc_v3f64_to_v3f32(ptr addrspace(1) %out, <3 x double> %in) {
 664 ; SI-LABEL: fptrunc_v3f64_to_v3f32:
 665 ; SI:       ; %bb.0:
 666 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
 667 ; SI-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x11
 668 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x15
 669 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 670 ; SI-NEXT:    s_mov_b32 s2, -1
 671 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 672 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 673 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 674 ; SI-NEXT:    v_cvt_f32_f64_e32 v2, s[4:5]
 675 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0 offset:8
 676 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 677 ; SI-NEXT:    s_endpgm
 678 ;
 679 ; VI-SDAG-LABEL: fptrunc_v3f64_to_v3f32:
 680 ; VI-SDAG:       ; %bb.0:
 681 ; VI-SDAG-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x54
 682 ; VI-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x44
 683 ; VI-SDAG-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x24
 684 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 685 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[6:7]
 686 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 687 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 688 ; VI-SDAG-NEXT:    s_mov_b32 s7, 0xf000
 689 ; VI-SDAG-NEXT:    s_mov_b32 s6, -1
 690 ; VI-SDAG-NEXT:    buffer_store_dwordx3 v[0:2], off, s[4:7], 0
 691 ; VI-SDAG-NEXT:    s_endpgm
 692 ;
 693 ; VI-GISEL-LABEL: fptrunc_v3f64_to_v3f32:
 694 ; VI-GISEL:       ; %bb.0:
 695 ; VI-GISEL-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
 696 ; VI-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 697 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
 698 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 699 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 700 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 701 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 702 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 703 ; VI-GISEL-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
 704 ; VI-GISEL-NEXT:    s_endpgm
 705 ;
 706 ; GFX10-SDAG-LABEL: fptrunc_v3f64_to_v3f32:
 707 ; GFX10-SDAG:       ; %bb.0:
 708 ; GFX10-SDAG-NEXT:    s_clause 0x1
 709 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x54
 710 ; GFX10-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x44
 711 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 712 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[6:7]
 713 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 714 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 715 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 716 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 717 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 718 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 719 ; GFX10-SDAG-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
 720 ; GFX10-SDAG-NEXT:    s_endpgm
 721 ;
 722 ; GFX10-GISEL-LABEL: fptrunc_v3f64_to_v3f32:
 723 ; GFX10-GISEL:       ; %bb.0:
 724 ; GFX10-GISEL-NEXT:    s_clause 0x1
 725 ; GFX10-GISEL-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
 726 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 727 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 728 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 729 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 730 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 731 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 732 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 733 ; GFX10-GISEL-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
 734 ; GFX10-GISEL-NEXT:    s_endpgm
 735 ;
 736 ; GFX11-SDAG-LABEL: fptrunc_v3f64_to_v3f32:
 737 ; GFX11-SDAG:       ; %bb.0:
 738 ; GFX11-SDAG-NEXT:    s_clause 0x1
 739 ; GFX11-SDAG-NEXT:    s_load_b64 s[6:7], s[4:5], 0x54
 740 ; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x44
 741 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 742 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[6:7]
 743 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[2:3]
 744 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[0:1]
 745 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
 746 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 747 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
 748 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 749 ; GFX11-SDAG-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], 0
 750 ; GFX11-SDAG-NEXT:    s_endpgm
 751 ;
 752 ; GFX11-GISEL-LABEL: fptrunc_v3f64_to_v3f32:
 753 ; GFX11-GISEL:       ; %bb.0:
 754 ; GFX11-GISEL-NEXT:    s_clause 0x1
 755 ; GFX11-GISEL-NEXT:    s_load_b256 s[8:15], s[4:5], 0x44
 756 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
 757 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
 758 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 759 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 760 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 761 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 762 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 763 ; GFX11-GISEL-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], 0
 764 ; GFX11-GISEL-NEXT:    s_endpgm
 765   %result = fptrunc <3 x double> %in to <3 x float>
 766   store <3 x float> %result, ptr addrspace(1) %out
 767   ret void
 768 }
 769
 770 define amdgpu_kernel void @fptrunc_v4f64_to_v4f32(ptr addrspace(1) %out, <4 x double> %in) {
 771 ; SI-LABEL: fptrunc_v4f64_to_v4f32:
 772 ; SI:       ; %bb.0:
 773 ; SI-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x11
 774 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
 775 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 776 ; SI-NEXT:    s_mov_b32 s2, -1
 777 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 778 ; SI-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 779 ; SI-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 780 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 781 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 782 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 783 ; SI-NEXT:    s_endpgm
 784 ;
 785 ; VI-SDAG-LABEL: fptrunc_v4f64_to_v4f32:
 786 ; VI-SDAG:       ; %bb.0:
 787 ; VI-SDAG-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
 788 ; VI-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 789 ; VI-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 790 ; VI-SDAG-NEXT:    s_mov_b32 s2, -1
 791 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 792 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 793 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 794 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 795 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 796 ; VI-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 797 ; VI-SDAG-NEXT:    s_endpgm
 798 ;
 799 ; VI-GISEL-LABEL: fptrunc_v4f64_to_v4f32:
 800 ; VI-GISEL:       ; %bb.0:
 801 ; VI-GISEL-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
 802 ; VI-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 803 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
 804 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 805 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 806 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 807 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 808 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 809 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 810 ; VI-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 811 ; VI-GISEL-NEXT:    s_endpgm
 812 ;
 813 ; GFX10-SDAG-LABEL: fptrunc_v4f64_to_v4f32:
 814 ; GFX10-SDAG:       ; %bb.0:
 815 ; GFX10-SDAG-NEXT:    s_clause 0x1
 816 ; GFX10-SDAG-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
 817 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 818 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 819 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 820 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 821 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 822 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 823 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 824 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 825 ; GFX10-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 826 ; GFX10-SDAG-NEXT:    s_endpgm
 827 ;
 828 ; GFX10-GISEL-LABEL: fptrunc_v4f64_to_v4f32:
 829 ; GFX10-GISEL:       ; %bb.0:
 830 ; GFX10-GISEL-NEXT:    s_clause 0x1
 831 ; GFX10-GISEL-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
 832 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 833 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 834 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 835 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 836 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 837 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 838 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 839 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 840 ; GFX10-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 841 ; GFX10-GISEL-NEXT:    s_endpgm
 842 ;
 843 ; GFX11-SDAG-LABEL: fptrunc_v4f64_to_v4f32:
 844 ; GFX11-SDAG:       ; %bb.0:
 845 ; GFX11-SDAG-NEXT:    s_clause 0x1
 846 ; GFX11-SDAG-NEXT:    s_load_b256 s[8:15], s[4:5], 0x44
 847 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
 848 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 849 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
 850 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 851 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 852 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 853 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 854 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 855 ; GFX11-SDAG-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
 856 ; GFX11-SDAG-NEXT:    s_endpgm
 857 ;
 858 ; GFX11-GISEL-LABEL: fptrunc_v4f64_to_v4f32:
 859 ; GFX11-GISEL:       ; %bb.0:
 860 ; GFX11-GISEL-NEXT:    s_clause 0x1
 861 ; GFX11-GISEL-NEXT:    s_load_b256 s[8:15], s[4:5], 0x44
 862 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
 863 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
 864 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 865 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 866 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 867 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 868 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 869 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 870 ; GFX11-GISEL-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
 871 ; GFX11-GISEL-NEXT:    s_endpgm
 872   %result = fptrunc <4 x double> %in to <4 x float>
 873   store <4 x float> %result, ptr addrspace(1) %out
 874   ret void
 875 }
 876
 877 define amdgpu_kernel void @fptrunc_v8f64_to_v8f32(ptr addrspace(1) %out, <8 x double> %in) {
 878 ; SI-LABEL: fptrunc_v8f64_to_v8f32:
 879 ; SI:       ; %bb.0:
 880 ; SI-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x19
 881 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
 882 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 883 ; SI-NEXT:    s_mov_b32 s2, -1
 884 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 885 ; SI-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 886 ; SI-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 887 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 888 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 889 ; SI-NEXT:    v_cvt_f32_f64_e32 v7, s[22:23]
 890 ; SI-NEXT:    v_cvt_f32_f64_e32 v6, s[20:21]
 891 ; SI-NEXT:    v_cvt_f32_f64_e32 v5, s[18:19]
 892 ; SI-NEXT:    v_cvt_f32_f64_e32 v4, s[16:17]
 893 ; SI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 894 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 895 ; SI-NEXT:    s_endpgm
 896 ;
 897 ; VI-SDAG-LABEL: fptrunc_v8f64_to_v8f32:
 898 ; VI-SDAG:       ; %bb.0:
 899 ; VI-SDAG-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x64
 900 ; VI-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 901 ; VI-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 902 ; VI-SDAG-NEXT:    s_mov_b32 s2, -1
 903 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 904 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v7, s[22:23]
 905 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v6, s[20:21]
 906 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v5, s[18:19]
 907 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v4, s[16:17]
 908 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 909 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 910 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 911 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 912 ; VI-SDAG-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 913 ; VI-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 914 ; VI-SDAG-NEXT:    s_endpgm
 915 ;
 916 ; VI-GISEL-LABEL: fptrunc_v8f64_to_v8f32:
 917 ; VI-GISEL:       ; %bb.0:
 918 ; VI-GISEL-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x64
 919 ; VI-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 920 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
 921 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 922 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 923 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 924 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 925 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 926 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 927 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v4, s[16:17]
 928 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v5, s[18:19]
 929 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v6, s[20:21]
 930 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v7, s[22:23]
 931 ; VI-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 932 ; VI-GISEL-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 933 ; VI-GISEL-NEXT:    s_endpgm
 934 ;
 935 ; GFX10-SDAG-LABEL: fptrunc_v8f64_to_v8f32:
 936 ; GFX10-SDAG:       ; %bb.0:
 937 ; GFX10-SDAG-NEXT:    s_clause 0x1
 938 ; GFX10-SDAG-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x64
 939 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 940 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 941 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 942 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 943 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v7, s[22:23]
 944 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v6, s[20:21]
 945 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v5, s[18:19]
 946 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v4, s[16:17]
 947 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 948 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 949 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 950 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 951 ; GFX10-SDAG-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 952 ; GFX10-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 953 ; GFX10-SDAG-NEXT:    s_endpgm
 954 ;
 955 ; GFX10-GISEL-LABEL: fptrunc_v8f64_to_v8f32:
 956 ; GFX10-GISEL:       ; %bb.0:
 957 ; GFX10-GISEL-NEXT:    s_clause 0x1
 958 ; GFX10-GISEL-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x64
 959 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 960 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 961 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 962 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 963 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 964 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 965 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 966 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 967 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v4, s[16:17]
 968 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v5, s[18:19]
 969 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v6, s[20:21]
 970 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v7, s[22:23]
 971 ; GFX10-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 972 ; GFX10-GISEL-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 973 ; GFX10-GISEL-NEXT:    s_endpgm
 974 ;
 975 ; GFX11-SDAG-LABEL: fptrunc_v8f64_to_v8f32:
 976 ; GFX11-SDAG:       ; %bb.0:
 977 ; GFX11-SDAG-NEXT:    s_clause 0x1
 978 ; GFX11-SDAG-NEXT:    s_load_b512 s[8:23], s[4:5], 0x64
 979 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
 980 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 981 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
 982 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 983 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v7, s[22:23]
 984 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v6, s[20:21]
 985 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v5, s[18:19]
 986 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v4, s[16:17]
 987 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
 988 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
 989 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 990 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 991 ; GFX11-SDAG-NEXT:    s_clause 0x1
 992 ; GFX11-SDAG-NEXT:    buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16
 993 ; GFX11-SDAG-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
 994 ; GFX11-SDAG-NEXT:    s_endpgm
 995 ;
 996 ; GFX11-GISEL-LABEL: fptrunc_v8f64_to_v8f32:
 997 ; GFX11-GISEL:       ; %bb.0:
 998 ; GFX11-GISEL-NEXT:    s_clause 0x1
 999 ; GFX11-GISEL-NEXT:    s_load_b512 s[8:23], s[4:5], 0x64
1000 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
1001 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
1002 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
1003 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
1004 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
1005 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
1006 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[12:13]
1007 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[14:15]
1008 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v4, s[16:17]
1009 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v5, s[18:19]
1010 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v6, s[20:21]
1011 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v7, s[22:23]
1012 ; GFX11-GISEL-NEXT:    s_clause 0x1
1013 ; GFX11-GISEL-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
1014 ; GFX11-GISEL-NEXT:    buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16
1015 ; GFX11-GISEL-NEXT:    s_endpgm
1016   %result = fptrunc <8 x double> %in to <8 x float>
1017   store <8 x float> %result, ptr addrspace(1) %out
1018   ret void
1019 }