llvm/test/CodeGen/AMDGPU/llvm.fmuladd.f16.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
   2 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math=preserve-sign -denormal-fp-math-f32=preserve-sign -verify-machineinstrs < %s | FileCheck --check-prefixes=SI %s
   3 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=fiji -denormal-fp-math=preserve-sign -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck --check-prefixes=VI-FLUSH %s
   4 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math=ieee -denormal-fp-math-f32=preserve-sign -verify-machineinstrs < %s | FileCheck --check-prefixes=SI %s
   5 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=fiji -denormal-fp-math=ieee -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck --check-prefixes=VI-DENORM %s
   6 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1010 -denormal-fp-math=preserve-sign -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-FLUSH %s
   7 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1010 -denormal-fp-math=ieee -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-DENORM %s
   8 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1100 -denormal-fp-math=preserve-sign -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11-FLUSH %s
   9 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1100 -denormal-fp-math=ieee -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11-DENORM %s
  10
  11 declare half @llvm.fmuladd.f16(half %a, half %b, half %c)
  12 declare <2 x half> @llvm.fmuladd.v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c)
  13
  14 define amdgpu_kernel void @fmuladd_f16(
  15 ; SI-LABEL: fmuladd_f16:
  16 ; SI:       ; %bb.0:
  17 ; SI-NEXT:    s_load_dwordx8 s[0:7], s[4:5], 0x9
  18 ; SI-NEXT:    s_mov_b32 s11, 0xf000
  19 ; SI-NEXT:    s_mov_b32 s10, -1
  20 ; SI-NEXT:    s_mov_b32 s14, s10
  21 ; SI-NEXT:    s_mov_b32 s15, s11
  22 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
  23 ; SI-NEXT:    s_mov_b32 s12, s2
  24 ; SI-NEXT:    s_mov_b32 s13, s3
  25 ; SI-NEXT:    s_mov_b32 s16, s4
  26 ; SI-NEXT:    s_mov_b32 s17, s5
  27 ; SI-NEXT:    s_mov_b32 s18, s10
  28 ; SI-NEXT:    s_mov_b32 s19, s11
  29 ; SI-NEXT:    s_mov_b32 s4, s6
  30 ; SI-NEXT:    s_mov_b32 s5, s7
  31 ; SI-NEXT:    s_mov_b32 s6, s10
  32 ; SI-NEXT:    s_mov_b32 s7, s11
  33 ; SI-NEXT:    buffer_load_ushort v0, off, s[12:15], 0
  34 ; SI-NEXT:    buffer_load_ushort v1, off, s[16:19], 0
  35 ; SI-NEXT:    buffer_load_ushort v2, off, s[4:7], 0
  36 ; SI-NEXT:    s_mov_b32 s8, s0
  37 ; SI-NEXT:    s_mov_b32 s9, s1
  38 ; SI-NEXT:    s_waitcnt vmcnt(2)
  39 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
  40 ; SI-NEXT:    s_waitcnt vmcnt(1)
  41 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
  42 ; SI-NEXT:    s_waitcnt vmcnt(0)
  43 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
  44 ; SI-NEXT:    v_mac_f32_e32 v2, v0, v1
  45 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v2
  46 ; SI-NEXT:    buffer_store_short v0, off, s[8:11], 0
  47 ; SI-NEXT:    s_endpgm
  48 ;
  49 ; VI-FLUSH-LABEL: fmuladd_f16:
  50 ; VI-FLUSH:       ; %bb.0:
  51 ; VI-FLUSH-NEXT:    s_load_dwordx8 s[0:7], s[4:5], 0x24
  52 ; VI-FLUSH-NEXT:    s_mov_b32 s11, 0xf000
  53 ; VI-FLUSH-NEXT:    s_mov_b32 s10, -1
  54 ; VI-FLUSH-NEXT:    s_mov_b32 s14, s10
  55 ; VI-FLUSH-NEXT:    s_mov_b32 s15, s11
  56 ; VI-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
  57 ; VI-FLUSH-NEXT:    s_mov_b32 s12, s2
  58 ; VI-FLUSH-NEXT:    s_mov_b32 s13, s3
  59 ; VI-FLUSH-NEXT:    s_mov_b32 s16, s4
  60 ; VI-FLUSH-NEXT:    s_mov_b32 s17, s5
  61 ; VI-FLUSH-NEXT:    s_mov_b32 s18, s10
  62 ; VI-FLUSH-NEXT:    s_mov_b32 s19, s11
  63 ; VI-FLUSH-NEXT:    s_mov_b32 s4, s6
  64 ; VI-FLUSH-NEXT:    s_mov_b32 s5, s7
  65 ; VI-FLUSH-NEXT:    s_mov_b32 s6, s10
  66 ; VI-FLUSH-NEXT:    s_mov_b32 s7, s11
  67 ; VI-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0
  68 ; VI-FLUSH-NEXT:    buffer_load_ushort v1, off, s[16:19], 0
  69 ; VI-FLUSH-NEXT:    buffer_load_ushort v2, off, s[4:7], 0
  70 ; VI-FLUSH-NEXT:    s_mov_b32 s8, s0
  71 ; VI-FLUSH-NEXT:    s_mov_b32 s9, s1
  72 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
  73 ; VI-FLUSH-NEXT:    v_mac_f16_e32 v2, v0, v1
  74 ; VI-FLUSH-NEXT:    buffer_store_short v2, off, s[8:11], 0
  75 ; VI-FLUSH-NEXT:    s_endpgm
  76 ;
  77 ; VI-DENORM-LABEL: fmuladd_f16:
  78 ; VI-DENORM:       ; %bb.0:
  79 ; VI-DENORM-NEXT:    s_load_dwordx8 s[0:7], s[4:5], 0x24
  80 ; VI-DENORM-NEXT:    s_mov_b32 s11, 0xf000
  81 ; VI-DENORM-NEXT:    s_mov_b32 s10, -1
  82 ; VI-DENORM-NEXT:    s_mov_b32 s14, s10
  83 ; VI-DENORM-NEXT:    s_mov_b32 s15, s11
  84 ; VI-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
  85 ; VI-DENORM-NEXT:    s_mov_b32 s12, s2
  86 ; VI-DENORM-NEXT:    s_mov_b32 s13, s3
  87 ; VI-DENORM-NEXT:    s_mov_b32 s16, s4
  88 ; VI-DENORM-NEXT:    s_mov_b32 s17, s5
  89 ; VI-DENORM-NEXT:    s_mov_b32 s18, s10
  90 ; VI-DENORM-NEXT:    s_mov_b32 s19, s11
  91 ; VI-DENORM-NEXT:    s_mov_b32 s4, s6
  92 ; VI-DENORM-NEXT:    s_mov_b32 s5, s7
  93 ; VI-DENORM-NEXT:    s_mov_b32 s6, s10
  94 ; VI-DENORM-NEXT:    s_mov_b32 s7, s11
  95 ; VI-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0
  96 ; VI-DENORM-NEXT:    buffer_load_ushort v1, off, s[16:19], 0
  97 ; VI-DENORM-NEXT:    buffer_load_ushort v2, off, s[4:7], 0
  98 ; VI-DENORM-NEXT:    s_mov_b32 s8, s0
  99 ; VI-DENORM-NEXT:    s_mov_b32 s9, s1
 100 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 101 ; VI-DENORM-NEXT:    v_fma_f16 v0, v0, v1, v2
 102 ; VI-DENORM-NEXT:    buffer_store_short v0, off, s[8:11], 0
 103 ; VI-DENORM-NEXT:    s_endpgm
 104 ;
 105 ; GFX10-FLUSH-LABEL: fmuladd_f16:
 106 ; GFX10-FLUSH:       ; %bb.0:
 107 ; GFX10-FLUSH-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x24
 108 ; GFX10-FLUSH-NEXT:    s_mov_b32 s2, -1
 109 ; GFX10-FLUSH-NEXT:    s_mov_b32 s3, 0x31016000
 110 ; GFX10-FLUSH-NEXT:    s_mov_b32 s6, s2
 111 ; GFX10-FLUSH-NEXT:    s_mov_b32 s7, s3
 112 ; GFX10-FLUSH-NEXT:    s_mov_b32 s18, s2
 113 ; GFX10-FLUSH-NEXT:    s_mov_b32 s19, s3
 114 ; GFX10-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 115 ; GFX10-FLUSH-NEXT:    s_mov_b32 s4, s10
 116 ; GFX10-FLUSH-NEXT:    s_mov_b32 s5, s11
 117 ; GFX10-FLUSH-NEXT:    s_mov_b32 s16, s12
 118 ; GFX10-FLUSH-NEXT:    s_mov_b32 s17, s13
 119 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v0, off, s[4:7], 0
 120 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v1, off, s[16:19], 0
 121 ; GFX10-FLUSH-NEXT:    s_mov_b32 s12, s14
 122 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s15
 123 ; GFX10-FLUSH-NEXT:    s_mov_b32 s14, s2
 124 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, s3
 125 ; GFX10-FLUSH-NEXT:    s_mov_b32 s0, s8
 126 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v2, off, s[12:15], 0
 127 ; GFX10-FLUSH-NEXT:    s_mov_b32 s1, s9
 128 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 129 ; GFX10-FLUSH-NEXT:    v_mul_f16_e32 v0, v0, v1
 130 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 131 ; GFX10-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v2
 132 ; GFX10-FLUSH-NEXT:    buffer_store_short v0, off, s[0:3], 0
 133 ; GFX10-FLUSH-NEXT:    s_endpgm
 134 ;
 135 ; GFX10-DENORM-LABEL: fmuladd_f16:
 136 ; GFX10-DENORM:       ; %bb.0:
 137 ; GFX10-DENORM-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x24
 138 ; GFX10-DENORM-NEXT:    s_mov_b32 s2, -1
 139 ; GFX10-DENORM-NEXT:    s_mov_b32 s3, 0x31016000
 140 ; GFX10-DENORM-NEXT:    s_mov_b32 s6, s2
 141 ; GFX10-DENORM-NEXT:    s_mov_b32 s7, s3
 142 ; GFX10-DENORM-NEXT:    s_mov_b32 s18, s2
 143 ; GFX10-DENORM-NEXT:    s_mov_b32 s19, s3
 144 ; GFX10-DENORM-NEXT:    s_mov_b32 s22, s2
 145 ; GFX10-DENORM-NEXT:    s_mov_b32 s23, s3
 146 ; GFX10-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 147 ; GFX10-DENORM-NEXT:    s_mov_b32 s4, s10
 148 ; GFX10-DENORM-NEXT:    s_mov_b32 s5, s11
 149 ; GFX10-DENORM-NEXT:    s_mov_b32 s16, s12
 150 ; GFX10-DENORM-NEXT:    s_mov_b32 s17, s13
 151 ; GFX10-DENORM-NEXT:    s_mov_b32 s20, s14
 152 ; GFX10-DENORM-NEXT:    s_mov_b32 s21, s15
 153 ; GFX10-DENORM-NEXT:    buffer_load_ushort v0, off, s[4:7], 0
 154 ; GFX10-DENORM-NEXT:    buffer_load_ushort v1, off, s[16:19], 0
 155 ; GFX10-DENORM-NEXT:    buffer_load_ushort v2, off, s[20:23], 0
 156 ; GFX10-DENORM-NEXT:    s_mov_b32 s0, s8
 157 ; GFX10-DENORM-NEXT:    s_mov_b32 s1, s9
 158 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 159 ; GFX10-DENORM-NEXT:    v_fmac_f16_e32 v2, v0, v1
 160 ; GFX10-DENORM-NEXT:    buffer_store_short v2, off, s[0:3], 0
 161 ; GFX10-DENORM-NEXT:    s_endpgm
 162 ;
 163 ; GFX11-FLUSH-LABEL: fmuladd_f16:
 164 ; GFX11-FLUSH:       ; %bb.0:
 165 ; GFX11-FLUSH-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
 166 ; GFX11-FLUSH-NEXT:    s_mov_b32 s10, -1
 167 ; GFX11-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 168 ; GFX11-FLUSH-NEXT:    s_mov_b32 s14, s10
 169 ; GFX11-FLUSH-NEXT:    s_mov_b32 s15, s11
 170 ; GFX11-FLUSH-NEXT:    s_mov_b32 s18, s10
 171 ; GFX11-FLUSH-NEXT:    s_mov_b32 s19, s11
 172 ; GFX11-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 173 ; GFX11-FLUSH-NEXT:    s_mov_b32 s12, s2
 174 ; GFX11-FLUSH-NEXT:    s_mov_b32 s13, s3
 175 ; GFX11-FLUSH-NEXT:    s_mov_b32 s16, s4
 176 ; GFX11-FLUSH-NEXT:    s_mov_b32 s17, s5
 177 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v0, off, s[12:15], 0
 178 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v1, off, s[16:19], 0
 179 ; GFX11-FLUSH-NEXT:    s_mov_b32 s4, s6
 180 ; GFX11-FLUSH-NEXT:    s_mov_b32 s5, s7
 181 ; GFX11-FLUSH-NEXT:    s_mov_b32 s6, s10
 182 ; GFX11-FLUSH-NEXT:    s_mov_b32 s7, s11
 183 ; GFX11-FLUSH-NEXT:    s_mov_b32 s8, s0
 184 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v2, off, s[4:7], 0
 185 ; GFX11-FLUSH-NEXT:    s_mov_b32 s9, s1
 186 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 187 ; GFX11-FLUSH-NEXT:    v_mul_f16_e32 v0, v0, v1
 188 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 189 ; GFX11-FLUSH-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 190 ; GFX11-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v2
 191 ; GFX11-FLUSH-NEXT:    buffer_store_b16 v0, off, s[8:11], 0
 192 ; GFX11-FLUSH-NEXT:    s_endpgm
 193 ;
 194 ; GFX11-DENORM-LABEL: fmuladd_f16:
 195 ; GFX11-DENORM:       ; %bb.0:
 196 ; GFX11-DENORM-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
 197 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 198 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 199 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 200 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 201 ; GFX11-DENORM-NEXT:    s_mov_b32 s18, s10
 202 ; GFX11-DENORM-NEXT:    s_mov_b32 s19, s11
 203 ; GFX11-DENORM-NEXT:    s_mov_b32 s22, s10
 204 ; GFX11-DENORM-NEXT:    s_mov_b32 s23, s11
 205 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 206 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s2
 207 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s3
 208 ; GFX11-DENORM-NEXT:    s_mov_b32 s16, s4
 209 ; GFX11-DENORM-NEXT:    s_mov_b32 s17, s5
 210 ; GFX11-DENORM-NEXT:    s_mov_b32 s20, s6
 211 ; GFX11-DENORM-NEXT:    s_mov_b32 s21, s7
 212 ; GFX11-DENORM-NEXT:    buffer_load_u16 v0, off, s[12:15], 0
 213 ; GFX11-DENORM-NEXT:    buffer_load_u16 v1, off, s[16:19], 0
 214 ; GFX11-DENORM-NEXT:    buffer_load_u16 v2, off, s[20:23], 0
 215 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s0
 216 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s1
 217 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 218 ; GFX11-DENORM-NEXT:    v_fmac_f16_e32 v2, v0, v1
 219 ; GFX11-DENORM-NEXT:    buffer_store_b16 v2, off, s[8:11], 0
 220 ; GFX11-DENORM-NEXT:    s_endpgm
 221     ptr addrspace(1) %r,
 222     ptr addrspace(1) %a,
 223     ptr addrspace(1) %b,
 224     ptr addrspace(1) %c) {
 225   %a.val = load half, ptr addrspace(1) %a
 226   %b.val = load half, ptr addrspace(1) %b
 227   %c.val = load half, ptr addrspace(1) %c
 228   %r.val = call half @llvm.fmuladd.f16(half %a.val, half %b.val, half %c.val)
 229   store half %r.val, ptr addrspace(1) %r
 230   ret void
 231 }
 232
 233 define amdgpu_kernel void @fmuladd_f16_imm_a(
 234 ; SI-LABEL: fmuladd_f16_imm_a:
 235 ; SI:       ; %bb.0:
 236 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
 237 ; SI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0xd
 238 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 239 ; SI-NEXT:    s_mov_b32 s6, -1
 240 ; SI-NEXT:    s_mov_b32 s14, s6
 241 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 242 ; SI-NEXT:    s_mov_b32 s12, s2
 243 ; SI-NEXT:    s_mov_b32 s13, s3
 244 ; SI-NEXT:    s_mov_b32 s15, s7
 245 ; SI-NEXT:    s_mov_b32 s10, s6
 246 ; SI-NEXT:    s_mov_b32 s11, s7
 247 ; SI-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 248 ; SI-NEXT:    s_waitcnt vmcnt(0)
 249 ; SI-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc
 250 ; SI-NEXT:    s_waitcnt vmcnt(0)
 251 ; SI-NEXT:    s_mov_b32 s4, s0
 252 ; SI-NEXT:    s_mov_b32 s5, s1
 253 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 254 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 255 ; SI-NEXT:    v_madmk_f32 v0, v0, 0x40400000, v1
 256 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 257 ; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
 258 ; SI-NEXT:    s_endpgm
 259 ;
 260 ; VI-FLUSH-LABEL: fmuladd_f16_imm_a:
 261 ; VI-FLUSH:       ; %bb.0:
 262 ; VI-FLUSH-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 263 ; VI-FLUSH-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 264 ; VI-FLUSH-NEXT:    s_mov_b32 s7, 0xf000
 265 ; VI-FLUSH-NEXT:    s_mov_b32 s6, -1
 266 ; VI-FLUSH-NEXT:    s_mov_b32 s14, s6
 267 ; VI-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 268 ; VI-FLUSH-NEXT:    s_mov_b32 s12, s2
 269 ; VI-FLUSH-NEXT:    s_mov_b32 s13, s3
 270 ; VI-FLUSH-NEXT:    s_mov_b32 s15, s7
 271 ; VI-FLUSH-NEXT:    s_mov_b32 s10, s6
 272 ; VI-FLUSH-NEXT:    s_mov_b32 s11, s7
 273 ; VI-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 274 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 275 ; VI-FLUSH-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc
 276 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 277 ; VI-FLUSH-NEXT:    s_mov_b32 s4, s0
 278 ; VI-FLUSH-NEXT:    s_mov_b32 s5, s1
 279 ; VI-FLUSH-NEXT:    v_madmk_f16 v0, v0, 0x4200, v1
 280 ; VI-FLUSH-NEXT:    buffer_store_short v0, off, s[4:7], 0
 281 ; VI-FLUSH-NEXT:    s_endpgm
 282 ;
 283 ; VI-DENORM-LABEL: fmuladd_f16_imm_a:
 284 ; VI-DENORM:       ; %bb.0:
 285 ; VI-DENORM-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 286 ; VI-DENORM-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 287 ; VI-DENORM-NEXT:    s_mov_b32 s7, 0xf000
 288 ; VI-DENORM-NEXT:    s_mov_b32 s6, -1
 289 ; VI-DENORM-NEXT:    s_mov_b32 s14, s6
 290 ; VI-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 291 ; VI-DENORM-NEXT:    s_mov_b32 s12, s2
 292 ; VI-DENORM-NEXT:    s_mov_b32 s13, s3
 293 ; VI-DENORM-NEXT:    s_mov_b32 s15, s7
 294 ; VI-DENORM-NEXT:    s_mov_b32 s10, s6
 295 ; VI-DENORM-NEXT:    s_mov_b32 s11, s7
 296 ; VI-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 297 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 298 ; VI-DENORM-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc
 299 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 300 ; VI-DENORM-NEXT:    s_mov_b32 s4, s0
 301 ; VI-DENORM-NEXT:    s_movk_i32 s0, 0x4200
 302 ; VI-DENORM-NEXT:    s_mov_b32 s5, s1
 303 ; VI-DENORM-NEXT:    v_fma_f16 v0, v0, s0, v1
 304 ; VI-DENORM-NEXT:    buffer_store_short v0, off, s[4:7], 0
 305 ; VI-DENORM-NEXT:    s_endpgm
 306 ;
 307 ; GFX10-FLUSH-LABEL: fmuladd_f16_imm_a:
 308 ; GFX10-FLUSH:       ; %bb.0:
 309 ; GFX10-FLUSH-NEXT:    s_clause 0x1
 310 ; GFX10-FLUSH-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 311 ; GFX10-FLUSH-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 312 ; GFX10-FLUSH-NEXT:    s_mov_b32 s6, -1
 313 ; GFX10-FLUSH-NEXT:    s_mov_b32 s7, 0x31016000
 314 ; GFX10-FLUSH-NEXT:    s_mov_b32 s14, s6
 315 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, s7
 316 ; GFX10-FLUSH-NEXT:    s_mov_b32 s10, s6
 317 ; GFX10-FLUSH-NEXT:    s_mov_b32 s11, s7
 318 ; GFX10-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 319 ; GFX10-FLUSH-NEXT:    s_mov_b32 s12, s2
 320 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s3
 321 ; GFX10-FLUSH-NEXT:    s_mov_b32 s4, s0
 322 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 323 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 324 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc dlc
 325 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 326 ; GFX10-FLUSH-NEXT:    s_mov_b32 s5, s1
 327 ; GFX10-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 328 ; GFX10-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 329 ; GFX10-FLUSH-NEXT:    buffer_store_short v0, off, s[4:7], 0
 330 ; GFX10-FLUSH-NEXT:    s_endpgm
 331 ;
 332 ; GFX10-DENORM-LABEL: fmuladd_f16_imm_a:
 333 ; GFX10-DENORM:       ; %bb.0:
 334 ; GFX10-DENORM-NEXT:    s_clause 0x1
 335 ; GFX10-DENORM-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 336 ; GFX10-DENORM-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 337 ; GFX10-DENORM-NEXT:    s_mov_b32 s6, -1
 338 ; GFX10-DENORM-NEXT:    s_mov_b32 s7, 0x31016000
 339 ; GFX10-DENORM-NEXT:    s_mov_b32 s14, s6
 340 ; GFX10-DENORM-NEXT:    s_mov_b32 s15, s7
 341 ; GFX10-DENORM-NEXT:    s_mov_b32 s10, s6
 342 ; GFX10-DENORM-NEXT:    s_mov_b32 s11, s7
 343 ; GFX10-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 344 ; GFX10-DENORM-NEXT:    s_mov_b32 s12, s2
 345 ; GFX10-DENORM-NEXT:    s_mov_b32 s13, s3
 346 ; GFX10-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 347 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 348 ; GFX10-DENORM-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc dlc
 349 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 350 ; GFX10-DENORM-NEXT:    s_mov_b32 s4, s0
 351 ; GFX10-DENORM-NEXT:    s_mov_b32 s5, s1
 352 ; GFX10-DENORM-NEXT:    v_fmamk_f16 v0, v0, 0x4200, v1
 353 ; GFX10-DENORM-NEXT:    buffer_store_short v0, off, s[4:7], 0
 354 ; GFX10-DENORM-NEXT:    s_endpgm
 355 ;
 356 ; GFX11-FLUSH-LABEL: fmuladd_f16_imm_a:
 357 ; GFX11-FLUSH:       ; %bb.0:
 358 ; GFX11-FLUSH-NEXT:    s_clause 0x1
 359 ; GFX11-FLUSH-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 360 ; GFX11-FLUSH-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
 361 ; GFX11-FLUSH-NEXT:    s_mov_b32 s10, -1
 362 ; GFX11-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 363 ; GFX11-FLUSH-NEXT:    s_mov_b32 s14, s10
 364 ; GFX11-FLUSH-NEXT:    s_mov_b32 s15, s11
 365 ; GFX11-FLUSH-NEXT:    s_mov_b32 s6, s10
 366 ; GFX11-FLUSH-NEXT:    s_mov_b32 s7, s11
 367 ; GFX11-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 368 ; GFX11-FLUSH-NEXT:    s_mov_b32 s12, s2
 369 ; GFX11-FLUSH-NEXT:    s_mov_b32 s13, s3
 370 ; GFX11-FLUSH-NEXT:    s_mov_b32 s8, s0
 371 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 372 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 373 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v1, off, s[4:7], 0 glc dlc
 374 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 375 ; GFX11-FLUSH-NEXT:    s_mov_b32 s9, s1
 376 ; GFX11-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 377 ; GFX11-FLUSH-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 378 ; GFX11-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 379 ; GFX11-FLUSH-NEXT:    buffer_store_b16 v0, off, s[8:11], 0
 380 ; GFX11-FLUSH-NEXT:    s_endpgm
 381 ;
 382 ; GFX11-DENORM-LABEL: fmuladd_f16_imm_a:
 383 ; GFX11-DENORM:       ; %bb.0:
 384 ; GFX11-DENORM-NEXT:    s_clause 0x1
 385 ; GFX11-DENORM-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 386 ; GFX11-DENORM-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
 387 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 388 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 389 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 390 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 391 ; GFX11-DENORM-NEXT:    s_mov_b32 s6, s10
 392 ; GFX11-DENORM-NEXT:    s_mov_b32 s7, s11
 393 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 394 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s2
 395 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s3
 396 ; GFX11-DENORM-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 397 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 398 ; GFX11-DENORM-NEXT:    buffer_load_u16 v1, off, s[4:7], 0 glc dlc
 399 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 400 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s0
 401 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s1
 402 ; GFX11-DENORM-NEXT:    v_fmac_f16_e32 v1, 0x4200, v0
 403 ; GFX11-DENORM-NEXT:    buffer_store_b16 v1, off, s[8:11], 0
 404 ; GFX11-DENORM-NEXT:    s_endpgm
 405     ptr addrspace(1) %r,
 406     ptr addrspace(1) %b,
 407     ptr addrspace(1) %c) {
 408   %b.val = load volatile half, ptr addrspace(1) %b
 409   %c.val = load volatile half, ptr addrspace(1) %c
 410   %r.val = call half @llvm.fmuladd.f16(half 3.0, half %b.val, half %c.val)
 411   store half %r.val, ptr addrspace(1) %r
 412   ret void
 413 }
 414
 415 define amdgpu_kernel void @fmuladd_f16_imm_b(
 416 ; SI-LABEL: fmuladd_f16_imm_b:
 417 ; SI:       ; %bb.0:
 418 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
 419 ; SI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0xd
 420 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 421 ; SI-NEXT:    s_mov_b32 s6, -1
 422 ; SI-NEXT:    s_mov_b32 s14, s6
 423 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 424 ; SI-NEXT:    s_mov_b32 s12, s2
 425 ; SI-NEXT:    s_mov_b32 s13, s3
 426 ; SI-NEXT:    s_mov_b32 s15, s7
 427 ; SI-NEXT:    s_mov_b32 s10, s6
 428 ; SI-NEXT:    s_mov_b32 s11, s7
 429 ; SI-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 430 ; SI-NEXT:    s_waitcnt vmcnt(0)
 431 ; SI-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc
 432 ; SI-NEXT:    s_waitcnt vmcnt(0)
 433 ; SI-NEXT:    s_mov_b32 s4, s0
 434 ; SI-NEXT:    s_mov_b32 s5, s1
 435 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 436 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 437 ; SI-NEXT:    v_madmk_f32 v0, v0, 0x40400000, v1
 438 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 439 ; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
 440 ; SI-NEXT:    s_endpgm
 441 ;
 442 ; VI-FLUSH-LABEL: fmuladd_f16_imm_b:
 443 ; VI-FLUSH:       ; %bb.0:
 444 ; VI-FLUSH-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 445 ; VI-FLUSH-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 446 ; VI-FLUSH-NEXT:    s_mov_b32 s7, 0xf000
 447 ; VI-FLUSH-NEXT:    s_mov_b32 s6, -1
 448 ; VI-FLUSH-NEXT:    s_mov_b32 s14, s6
 449 ; VI-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 450 ; VI-FLUSH-NEXT:    s_mov_b32 s12, s2
 451 ; VI-FLUSH-NEXT:    s_mov_b32 s13, s3
 452 ; VI-FLUSH-NEXT:    s_mov_b32 s15, s7
 453 ; VI-FLUSH-NEXT:    s_mov_b32 s10, s6
 454 ; VI-FLUSH-NEXT:    s_mov_b32 s11, s7
 455 ; VI-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 456 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 457 ; VI-FLUSH-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc
 458 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 459 ; VI-FLUSH-NEXT:    s_mov_b32 s4, s0
 460 ; VI-FLUSH-NEXT:    s_mov_b32 s5, s1
 461 ; VI-FLUSH-NEXT:    v_madmk_f16 v0, v0, 0x4200, v1
 462 ; VI-FLUSH-NEXT:    buffer_store_short v0, off, s[4:7], 0
 463 ; VI-FLUSH-NEXT:    s_endpgm
 464 ;
 465 ; VI-DENORM-LABEL: fmuladd_f16_imm_b:
 466 ; VI-DENORM:       ; %bb.0:
 467 ; VI-DENORM-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 468 ; VI-DENORM-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 469 ; VI-DENORM-NEXT:    s_mov_b32 s7, 0xf000
 470 ; VI-DENORM-NEXT:    s_mov_b32 s6, -1
 471 ; VI-DENORM-NEXT:    s_mov_b32 s14, s6
 472 ; VI-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 473 ; VI-DENORM-NEXT:    s_mov_b32 s12, s2
 474 ; VI-DENORM-NEXT:    s_mov_b32 s13, s3
 475 ; VI-DENORM-NEXT:    s_mov_b32 s15, s7
 476 ; VI-DENORM-NEXT:    s_mov_b32 s10, s6
 477 ; VI-DENORM-NEXT:    s_mov_b32 s11, s7
 478 ; VI-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 479 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 480 ; VI-DENORM-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc
 481 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 482 ; VI-DENORM-NEXT:    s_mov_b32 s4, s0
 483 ; VI-DENORM-NEXT:    s_movk_i32 s0, 0x4200
 484 ; VI-DENORM-NEXT:    s_mov_b32 s5, s1
 485 ; VI-DENORM-NEXT:    v_fma_f16 v0, v0, s0, v1
 486 ; VI-DENORM-NEXT:    buffer_store_short v0, off, s[4:7], 0
 487 ; VI-DENORM-NEXT:    s_endpgm
 488 ;
 489 ; GFX10-FLUSH-LABEL: fmuladd_f16_imm_b:
 490 ; GFX10-FLUSH:       ; %bb.0:
 491 ; GFX10-FLUSH-NEXT:    s_clause 0x1
 492 ; GFX10-FLUSH-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 493 ; GFX10-FLUSH-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 494 ; GFX10-FLUSH-NEXT:    s_mov_b32 s6, -1
 495 ; GFX10-FLUSH-NEXT:    s_mov_b32 s7, 0x31016000
 496 ; GFX10-FLUSH-NEXT:    s_mov_b32 s14, s6
 497 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, s7
 498 ; GFX10-FLUSH-NEXT:    s_mov_b32 s10, s6
 499 ; GFX10-FLUSH-NEXT:    s_mov_b32 s11, s7
 500 ; GFX10-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 501 ; GFX10-FLUSH-NEXT:    s_mov_b32 s12, s2
 502 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s3
 503 ; GFX10-FLUSH-NEXT:    s_mov_b32 s4, s0
 504 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 505 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 506 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc dlc
 507 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 508 ; GFX10-FLUSH-NEXT:    s_mov_b32 s5, s1
 509 ; GFX10-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 510 ; GFX10-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 511 ; GFX10-FLUSH-NEXT:    buffer_store_short v0, off, s[4:7], 0
 512 ; GFX10-FLUSH-NEXT:    s_endpgm
 513 ;
 514 ; GFX10-DENORM-LABEL: fmuladd_f16_imm_b:
 515 ; GFX10-DENORM:       ; %bb.0:
 516 ; GFX10-DENORM-NEXT:    s_clause 0x1
 517 ; GFX10-DENORM-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 518 ; GFX10-DENORM-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x34
 519 ; GFX10-DENORM-NEXT:    s_mov_b32 s6, -1
 520 ; GFX10-DENORM-NEXT:    s_mov_b32 s7, 0x31016000
 521 ; GFX10-DENORM-NEXT:    s_mov_b32 s14, s6
 522 ; GFX10-DENORM-NEXT:    s_mov_b32 s15, s7
 523 ; GFX10-DENORM-NEXT:    s_mov_b32 s10, s6
 524 ; GFX10-DENORM-NEXT:    s_mov_b32 s11, s7
 525 ; GFX10-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 526 ; GFX10-DENORM-NEXT:    s_mov_b32 s12, s2
 527 ; GFX10-DENORM-NEXT:    s_mov_b32 s13, s3
 528 ; GFX10-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 529 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 530 ; GFX10-DENORM-NEXT:    buffer_load_ushort v1, off, s[8:11], 0 glc dlc
 531 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 532 ; GFX10-DENORM-NEXT:    s_mov_b32 s4, s0
 533 ; GFX10-DENORM-NEXT:    s_mov_b32 s5, s1
 534 ; GFX10-DENORM-NEXT:    v_fmamk_f16 v0, v0, 0x4200, v1
 535 ; GFX10-DENORM-NEXT:    buffer_store_short v0, off, s[4:7], 0
 536 ; GFX10-DENORM-NEXT:    s_endpgm
 537 ;
 538 ; GFX11-FLUSH-LABEL: fmuladd_f16_imm_b:
 539 ; GFX11-FLUSH:       ; %bb.0:
 540 ; GFX11-FLUSH-NEXT:    s_clause 0x1
 541 ; GFX11-FLUSH-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 542 ; GFX11-FLUSH-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
 543 ; GFX11-FLUSH-NEXT:    s_mov_b32 s10, -1
 544 ; GFX11-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 545 ; GFX11-FLUSH-NEXT:    s_mov_b32 s14, s10
 546 ; GFX11-FLUSH-NEXT:    s_mov_b32 s15, s11
 547 ; GFX11-FLUSH-NEXT:    s_mov_b32 s6, s10
 548 ; GFX11-FLUSH-NEXT:    s_mov_b32 s7, s11
 549 ; GFX11-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 550 ; GFX11-FLUSH-NEXT:    s_mov_b32 s12, s2
 551 ; GFX11-FLUSH-NEXT:    s_mov_b32 s13, s3
 552 ; GFX11-FLUSH-NEXT:    s_mov_b32 s8, s0
 553 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 554 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 555 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v1, off, s[4:7], 0 glc dlc
 556 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 557 ; GFX11-FLUSH-NEXT:    s_mov_b32 s9, s1
 558 ; GFX11-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 559 ; GFX11-FLUSH-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 560 ; GFX11-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 561 ; GFX11-FLUSH-NEXT:    buffer_store_b16 v0, off, s[8:11], 0
 562 ; GFX11-FLUSH-NEXT:    s_endpgm
 563 ;
 564 ; GFX11-DENORM-LABEL: fmuladd_f16_imm_b:
 565 ; GFX11-DENORM:       ; %bb.0:
 566 ; GFX11-DENORM-NEXT:    s_clause 0x1
 567 ; GFX11-DENORM-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 568 ; GFX11-DENORM-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
 569 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 570 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 571 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 572 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 573 ; GFX11-DENORM-NEXT:    s_mov_b32 s6, s10
 574 ; GFX11-DENORM-NEXT:    s_mov_b32 s7, s11
 575 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 576 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s2
 577 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s3
 578 ; GFX11-DENORM-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 579 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 580 ; GFX11-DENORM-NEXT:    buffer_load_u16 v1, off, s[4:7], 0 glc dlc
 581 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 582 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s0
 583 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s1
 584 ; GFX11-DENORM-NEXT:    v_fmac_f16_e32 v1, 0x4200, v0
 585 ; GFX11-DENORM-NEXT:    buffer_store_b16 v1, off, s[8:11], 0
 586 ; GFX11-DENORM-NEXT:    s_endpgm
 587     ptr addrspace(1) %r,
 588     ptr addrspace(1) %a,
 589     ptr addrspace(1) %c) {
 590   %a.val = load volatile half, ptr addrspace(1) %a
 591   %c.val = load volatile half, ptr addrspace(1) %c
 592   %r.val = call half @llvm.fmuladd.f16(half %a.val, half 3.0, half %c.val)
 593   store half %r.val, ptr addrspace(1) %r
 594   ret void
 595 }
 596
 597 define amdgpu_kernel void @fmuladd_v2f16(
 598 ; SI-LABEL: fmuladd_v2f16:
 599 ; SI:       ; %bb.0:
 600 ; SI-NEXT:    s_load_dwordx8 s[0:7], s[4:5], 0x9
 601 ; SI-NEXT:    s_mov_b32 s11, 0xf000
 602 ; SI-NEXT:    s_mov_b32 s10, -1
 603 ; SI-NEXT:    s_mov_b32 s14, s10
 604 ; SI-NEXT:    s_mov_b32 s15, s11
 605 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 606 ; SI-NEXT:    s_mov_b32 s12, s2
 607 ; SI-NEXT:    s_mov_b32 s13, s3
 608 ; SI-NEXT:    s_mov_b32 s16, s4
 609 ; SI-NEXT:    s_mov_b32 s17, s5
 610 ; SI-NEXT:    s_mov_b32 s18, s10
 611 ; SI-NEXT:    s_mov_b32 s19, s11
 612 ; SI-NEXT:    s_mov_b32 s4, s6
 613 ; SI-NEXT:    s_mov_b32 s5, s7
 614 ; SI-NEXT:    s_mov_b32 s6, s10
 615 ; SI-NEXT:    buffer_load_dword v0, off, s[12:15], 0
 616 ; SI-NEXT:    s_mov_b32 s7, s11
 617 ; SI-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 618 ; SI-NEXT:    buffer_load_dword v2, off, s[4:7], 0
 619 ; SI-NEXT:    s_mov_b32 s8, s0
 620 ; SI-NEXT:    s_mov_b32 s9, s1
 621 ; SI-NEXT:    s_waitcnt vmcnt(2)
 622 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v0
 623 ; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
 624 ; SI-NEXT:    s_waitcnt vmcnt(1)
 625 ; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
 626 ; SI-NEXT:    s_waitcnt vmcnt(0)
 627 ; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
 628 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 629 ; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
 630 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
 631 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 632 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 633 ; SI-NEXT:    v_mac_f32_e32 v5, v0, v4
 634 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v5
 635 ; SI-NEXT:    v_mac_f32_e32 v2, v3, v1
 636 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v2
 637 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 638 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 639 ; SI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
 640 ; SI-NEXT:    s_endpgm
 641 ;
 642 ; VI-FLUSH-LABEL: fmuladd_v2f16:
 643 ; VI-FLUSH:       ; %bb.0:
 644 ; VI-FLUSH-NEXT:    s_load_dwordx8 s[0:7], s[4:5], 0x24
 645 ; VI-FLUSH-NEXT:    s_mov_b32 s11, 0xf000
 646 ; VI-FLUSH-NEXT:    s_mov_b32 s10, -1
 647 ; VI-FLUSH-NEXT:    s_mov_b32 s14, s10
 648 ; VI-FLUSH-NEXT:    s_mov_b32 s15, s11
 649 ; VI-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 650 ; VI-FLUSH-NEXT:    s_mov_b32 s12, s2
 651 ; VI-FLUSH-NEXT:    s_mov_b32 s13, s3
 652 ; VI-FLUSH-NEXT:    s_mov_b32 s16, s4
 653 ; VI-FLUSH-NEXT:    s_mov_b32 s17, s5
 654 ; VI-FLUSH-NEXT:    s_mov_b32 s18, s10
 655 ; VI-FLUSH-NEXT:    s_mov_b32 s19, s11
 656 ; VI-FLUSH-NEXT:    s_mov_b32 s4, s6
 657 ; VI-FLUSH-NEXT:    s_mov_b32 s5, s7
 658 ; VI-FLUSH-NEXT:    s_mov_b32 s6, s10
 659 ; VI-FLUSH-NEXT:    s_mov_b32 s7, s11
 660 ; VI-FLUSH-NEXT:    buffer_load_dword v0, off, s[12:15], 0
 661 ; VI-FLUSH-NEXT:    buffer_load_dword v1, off, s[4:7], 0
 662 ; VI-FLUSH-NEXT:    buffer_load_dword v2, off, s[16:19], 0
 663 ; VI-FLUSH-NEXT:    s_mov_b32 s8, s0
 664 ; VI-FLUSH-NEXT:    s_mov_b32 s9, s1
 665 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 666 ; VI-FLUSH-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
 667 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 668 ; VI-FLUSH-NEXT:    v_mac_f16_sdwa v3, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 669 ; VI-FLUSH-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 670 ; VI-FLUSH-NEXT:    v_mac_f16_e32 v1, v0, v2
 671 ; VI-FLUSH-NEXT:    v_or_b32_e32 v0, v1, v3
 672 ; VI-FLUSH-NEXT:    buffer_store_dword v0, off, s[8:11], 0
 673 ; VI-FLUSH-NEXT:    s_endpgm
 674 ;
 675 ; VI-DENORM-LABEL: fmuladd_v2f16:
 676 ; VI-DENORM:       ; %bb.0:
 677 ; VI-DENORM-NEXT:    s_load_dwordx8 s[0:7], s[4:5], 0x24
 678 ; VI-DENORM-NEXT:    s_mov_b32 s11, 0xf000
 679 ; VI-DENORM-NEXT:    s_mov_b32 s10, -1
 680 ; VI-DENORM-NEXT:    s_mov_b32 s14, s10
 681 ; VI-DENORM-NEXT:    s_mov_b32 s15, s11
 682 ; VI-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 683 ; VI-DENORM-NEXT:    s_mov_b32 s16, s4
 684 ; VI-DENORM-NEXT:    s_mov_b32 s17, s5
 685 ; VI-DENORM-NEXT:    s_mov_b32 s4, s6
 686 ; VI-DENORM-NEXT:    s_mov_b32 s5, s7
 687 ; VI-DENORM-NEXT:    s_mov_b32 s6, s10
 688 ; VI-DENORM-NEXT:    s_mov_b32 s7, s11
 689 ; VI-DENORM-NEXT:    s_mov_b32 s12, s2
 690 ; VI-DENORM-NEXT:    s_mov_b32 s13, s3
 691 ; VI-DENORM-NEXT:    s_mov_b32 s18, s10
 692 ; VI-DENORM-NEXT:    s_mov_b32 s19, s11
 693 ; VI-DENORM-NEXT:    buffer_load_dword v0, off, s[4:7], 0
 694 ; VI-DENORM-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 695 ; VI-DENORM-NEXT:    buffer_load_dword v2, off, s[12:15], 0
 696 ; VI-DENORM-NEXT:    s_mov_b32 s8, s0
 697 ; VI-DENORM-NEXT:    s_mov_b32 s9, s1
 698 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(2)
 699 ; VI-DENORM-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
 700 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(1)
 701 ; VI-DENORM-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
 702 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 703 ; VI-DENORM-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
 704 ; VI-DENORM-NEXT:    v_fma_f16 v3, v5, v4, v3
 705 ; VI-DENORM-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 706 ; VI-DENORM-NEXT:    v_fma_f16 v0, v2, v1, v0
 707 ; VI-DENORM-NEXT:    v_or_b32_e32 v0, v0, v3
 708 ; VI-DENORM-NEXT:    buffer_store_dword v0, off, s[8:11], 0
 709 ; VI-DENORM-NEXT:    s_endpgm
 710 ;
 711 ; GFX10-FLUSH-LABEL: fmuladd_v2f16:
 712 ; GFX10-FLUSH:       ; %bb.0:
 713 ; GFX10-FLUSH-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x24
 714 ; GFX10-FLUSH-NEXT:    s_mov_b32 s2, -1
 715 ; GFX10-FLUSH-NEXT:    s_mov_b32 s3, 0x31016000
 716 ; GFX10-FLUSH-NEXT:    s_mov_b32 s6, s2
 717 ; GFX10-FLUSH-NEXT:    s_mov_b32 s7, s3
 718 ; GFX10-FLUSH-NEXT:    s_mov_b32 s18, s2
 719 ; GFX10-FLUSH-NEXT:    s_mov_b32 s19, s3
 720 ; GFX10-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 721 ; GFX10-FLUSH-NEXT:    s_mov_b32 s4, s10
 722 ; GFX10-FLUSH-NEXT:    s_mov_b32 s5, s11
 723 ; GFX10-FLUSH-NEXT:    s_mov_b32 s16, s12
 724 ; GFX10-FLUSH-NEXT:    s_mov_b32 s17, s13
 725 ; GFX10-FLUSH-NEXT:    buffer_load_dword v0, off, s[4:7], 0
 726 ; GFX10-FLUSH-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 727 ; GFX10-FLUSH-NEXT:    s_mov_b32 s12, s14
 728 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s15
 729 ; GFX10-FLUSH-NEXT:    s_mov_b32 s14, s2
 730 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, s3
 731 ; GFX10-FLUSH-NEXT:    s_mov_b32 s0, s8
 732 ; GFX10-FLUSH-NEXT:    buffer_load_dword v2, off, s[12:15], 0
 733 ; GFX10-FLUSH-NEXT:    s_mov_b32 s1, s9
 734 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 735 ; GFX10-FLUSH-NEXT:    v_pk_mul_f16 v0, v0, v1
 736 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 737 ; GFX10-FLUSH-NEXT:    v_pk_add_f16 v0, v0, v2
 738 ; GFX10-FLUSH-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 739 ; GFX10-FLUSH-NEXT:    s_endpgm
 740 ;
 741 ; GFX10-DENORM-LABEL: fmuladd_v2f16:
 742 ; GFX10-DENORM:       ; %bb.0:
 743 ; GFX10-DENORM-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x24
 744 ; GFX10-DENORM-NEXT:    s_mov_b32 s2, -1
 745 ; GFX10-DENORM-NEXT:    s_mov_b32 s3, 0x31016000
 746 ; GFX10-DENORM-NEXT:    s_mov_b32 s6, s2
 747 ; GFX10-DENORM-NEXT:    s_mov_b32 s7, s3
 748 ; GFX10-DENORM-NEXT:    s_mov_b32 s18, s2
 749 ; GFX10-DENORM-NEXT:    s_mov_b32 s19, s3
 750 ; GFX10-DENORM-NEXT:    s_mov_b32 s22, s2
 751 ; GFX10-DENORM-NEXT:    s_mov_b32 s23, s3
 752 ; GFX10-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 753 ; GFX10-DENORM-NEXT:    s_mov_b32 s4, s10
 754 ; GFX10-DENORM-NEXT:    s_mov_b32 s5, s11
 755 ; GFX10-DENORM-NEXT:    s_mov_b32 s16, s12
 756 ; GFX10-DENORM-NEXT:    s_mov_b32 s17, s13
 757 ; GFX10-DENORM-NEXT:    s_mov_b32 s20, s14
 758 ; GFX10-DENORM-NEXT:    s_mov_b32 s21, s15
 759 ; GFX10-DENORM-NEXT:    buffer_load_dword v0, off, s[4:7], 0
 760 ; GFX10-DENORM-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 761 ; GFX10-DENORM-NEXT:    buffer_load_dword v2, off, s[20:23], 0
 762 ; GFX10-DENORM-NEXT:    s_mov_b32 s0, s8
 763 ; GFX10-DENORM-NEXT:    s_mov_b32 s1, s9
 764 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 765 ; GFX10-DENORM-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
 766 ; GFX10-DENORM-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 767 ; GFX10-DENORM-NEXT:    s_endpgm
 768 ;
 769 ; GFX11-FLUSH-LABEL: fmuladd_v2f16:
 770 ; GFX11-FLUSH:       ; %bb.0:
 771 ; GFX11-FLUSH-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
 772 ; GFX11-FLUSH-NEXT:    s_mov_b32 s10, -1
 773 ; GFX11-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 774 ; GFX11-FLUSH-NEXT:    s_mov_b32 s14, s10
 775 ; GFX11-FLUSH-NEXT:    s_mov_b32 s15, s11
 776 ; GFX11-FLUSH-NEXT:    s_mov_b32 s18, s10
 777 ; GFX11-FLUSH-NEXT:    s_mov_b32 s19, s11
 778 ; GFX11-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 779 ; GFX11-FLUSH-NEXT:    s_mov_b32 s12, s2
 780 ; GFX11-FLUSH-NEXT:    s_mov_b32 s13, s3
 781 ; GFX11-FLUSH-NEXT:    s_mov_b32 s16, s4
 782 ; GFX11-FLUSH-NEXT:    s_mov_b32 s17, s5
 783 ; GFX11-FLUSH-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
 784 ; GFX11-FLUSH-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
 785 ; GFX11-FLUSH-NEXT:    s_mov_b32 s4, s6
 786 ; GFX11-FLUSH-NEXT:    s_mov_b32 s5, s7
 787 ; GFX11-FLUSH-NEXT:    s_mov_b32 s6, s10
 788 ; GFX11-FLUSH-NEXT:    s_mov_b32 s7, s11
 789 ; GFX11-FLUSH-NEXT:    s_mov_b32 s8, s0
 790 ; GFX11-FLUSH-NEXT:    buffer_load_b32 v2, off, s[4:7], 0
 791 ; GFX11-FLUSH-NEXT:    s_mov_b32 s9, s1
 792 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 793 ; GFX11-FLUSH-NEXT:    v_pk_mul_f16 v0, v0, v1
 794 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 795 ; GFX11-FLUSH-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 796 ; GFX11-FLUSH-NEXT:    v_pk_add_f16 v0, v0, v2
 797 ; GFX11-FLUSH-NEXT:    buffer_store_b32 v0, off, s[8:11], 0
 798 ; GFX11-FLUSH-NEXT:    s_endpgm
 799 ;
 800 ; GFX11-DENORM-LABEL: fmuladd_v2f16:
 801 ; GFX11-DENORM:       ; %bb.0:
 802 ; GFX11-DENORM-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
 803 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 804 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 805 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 806 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 807 ; GFX11-DENORM-NEXT:    s_mov_b32 s18, s10
 808 ; GFX11-DENORM-NEXT:    s_mov_b32 s19, s11
 809 ; GFX11-DENORM-NEXT:    s_mov_b32 s22, s10
 810 ; GFX11-DENORM-NEXT:    s_mov_b32 s23, s11
 811 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 812 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s2
 813 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s3
 814 ; GFX11-DENORM-NEXT:    s_mov_b32 s16, s4
 815 ; GFX11-DENORM-NEXT:    s_mov_b32 s17, s5
 816 ; GFX11-DENORM-NEXT:    s_mov_b32 s20, s6
 817 ; GFX11-DENORM-NEXT:    s_mov_b32 s21, s7
 818 ; GFX11-DENORM-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
 819 ; GFX11-DENORM-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
 820 ; GFX11-DENORM-NEXT:    buffer_load_b32 v2, off, s[20:23], 0
 821 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s0
 822 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s1
 823 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 824 ; GFX11-DENORM-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
 825 ; GFX11-DENORM-NEXT:    buffer_store_b32 v0, off, s[8:11], 0
 826 ; GFX11-DENORM-NEXT:    s_endpgm
 827     ptr addrspace(1) %r,
 828     ptr addrspace(1) %a,
 829     ptr addrspace(1) %b,
 830     ptr addrspace(1) %c) {
 831   %a.val = load <2 x half>, ptr addrspace(1) %a
 832   %b.val = load <2 x half>, ptr addrspace(1) %b
 833   %c.val = load <2 x half>, ptr addrspace(1) %c
 834   %r.val = call <2 x half> @llvm.fmuladd.v2f16(<2 x half> %a.val, <2 x half> %b.val, <2 x half> %c.val)
 835   store <2 x half> %r.val, ptr addrspace(1) %r
 836   ret void
 837 }