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[2:3], 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[2:3], 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[2:3], 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[4:11], s[2:3], 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 s14, s2
 111 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, 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 s12, s6
 116 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s7
 117 ; GFX10-FLUSH-NEXT:    s_mov_b32 s16, s8
 118 ; GFX10-FLUSH-NEXT:    s_mov_b32 s17, s9
 119 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0
 120 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v1, off, s[16:19], 0
 121 ; GFX10-FLUSH-NEXT:    s_mov_b32 s8, s10
 122 ; GFX10-FLUSH-NEXT:    s_mov_b32 s9, s11
 123 ; GFX10-FLUSH-NEXT:    s_mov_b32 s10, s2
 124 ; GFX10-FLUSH-NEXT:    s_mov_b32 s11, s3
 125 ; GFX10-FLUSH-NEXT:    s_mov_b32 s0, s4
 126 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v2, off, s[8:11], 0
 127 ; GFX10-FLUSH-NEXT:    s_mov_b32 s1, s5
 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[4:11], s[2:3], 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 s14, s2
 141 ; GFX10-DENORM-NEXT:    s_mov_b32 s15, 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 s12, s6
 148 ; GFX10-DENORM-NEXT:    s_mov_b32 s13, s7
 149 ; GFX10-DENORM-NEXT:    s_mov_b32 s16, s8
 150 ; GFX10-DENORM-NEXT:    s_mov_b32 s17, s9
 151 ; GFX10-DENORM-NEXT:    s_mov_b32 s20, s10
 152 ; GFX10-DENORM-NEXT:    s_mov_b32 s21, s11
 153 ; GFX10-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 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, s4
 157 ; GFX10-DENORM-NEXT:    s_mov_b32 s1, s5
 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[2:3], 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_nop 0
 193 ; GFX11-FLUSH-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 194 ; GFX11-FLUSH-NEXT:    s_endpgm
 195 ;
 196 ; GFX11-DENORM-LABEL: fmuladd_f16:
 197 ; GFX11-DENORM:       ; %bb.0:
 198 ; GFX11-DENORM-NEXT:    s_load_b256 s[0:7], s[2:3], 0x24
 199 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 200 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 201 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 202 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 203 ; GFX11-DENORM-NEXT:    s_mov_b32 s18, s10
 204 ; GFX11-DENORM-NEXT:    s_mov_b32 s19, s11
 205 ; GFX11-DENORM-NEXT:    s_mov_b32 s22, s10
 206 ; GFX11-DENORM-NEXT:    s_mov_b32 s23, s11
 207 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 208 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s2
 209 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s3
 210 ; GFX11-DENORM-NEXT:    s_mov_b32 s16, s4
 211 ; GFX11-DENORM-NEXT:    s_mov_b32 s17, s5
 212 ; GFX11-DENORM-NEXT:    s_mov_b32 s20, s6
 213 ; GFX11-DENORM-NEXT:    s_mov_b32 s21, s7
 214 ; GFX11-DENORM-NEXT:    buffer_load_u16 v0, off, s[12:15], 0
 215 ; GFX11-DENORM-NEXT:    buffer_load_u16 v1, off, s[16:19], 0
 216 ; GFX11-DENORM-NEXT:    buffer_load_u16 v2, off, s[20:23], 0
 217 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s0
 218 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s1
 219 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 220 ; GFX11-DENORM-NEXT:    v_fmac_f16_e32 v2, v0, v1
 221 ; GFX11-DENORM-NEXT:    buffer_store_b16 v2, off, s[8:11], 0
 222 ; GFX11-DENORM-NEXT:    s_nop 0
 223 ; GFX11-DENORM-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 224 ; GFX11-DENORM-NEXT:    s_endpgm
 225     ptr addrspace(1) %r,
 226     ptr addrspace(1) %a,
 227     ptr addrspace(1) %b,
 228     ptr addrspace(1) %c) {
 229   %a.val = load half, ptr addrspace(1) %a
 230   %b.val = load half, ptr addrspace(1) %b
 231   %c.val = load half, ptr addrspace(1) %c
 232   %r.val = call half @llvm.fmuladd.f16(half %a.val, half %b.val, half %c.val)
 233   store half %r.val, ptr addrspace(1) %r
 234   ret void
 235 }
 236
 237 define amdgpu_kernel void @fmuladd_f16_imm_a(
 238 ; SI-LABEL: fmuladd_f16_imm_a:
 239 ; SI:       ; %bb.0:
 240 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
 241 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xd
 242 ; SI-NEXT:    s_mov_b32 s11, 0xf000
 243 ; SI-NEXT:    s_mov_b32 s10, -1
 244 ; SI-NEXT:    s_mov_b32 s14, s10
 245 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 246 ; SI-NEXT:    s_mov_b32 s12, s6
 247 ; SI-NEXT:    s_mov_b32 s13, s7
 248 ; SI-NEXT:    s_mov_b32 s15, s11
 249 ; SI-NEXT:    s_mov_b32 s2, s10
 250 ; SI-NEXT:    s_mov_b32 s3, s11
 251 ; SI-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 252 ; SI-NEXT:    s_waitcnt vmcnt(0)
 253 ; SI-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc
 254 ; SI-NEXT:    s_waitcnt vmcnt(0)
 255 ; SI-NEXT:    s_mov_b32 s8, s4
 256 ; SI-NEXT:    s_mov_b32 s9, s5
 257 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 258 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 259 ; SI-NEXT:    v_madmk_f32 v0, v0, 0x40400000, v1
 260 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 261 ; SI-NEXT:    buffer_store_short v0, off, s[8:11], 0
 262 ; SI-NEXT:    s_endpgm
 263 ;
 264 ; VI-FLUSH-LABEL: fmuladd_f16_imm_a:
 265 ; VI-FLUSH:       ; %bb.0:
 266 ; VI-FLUSH-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 267 ; VI-FLUSH-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 268 ; VI-FLUSH-NEXT:    s_mov_b32 s11, 0xf000
 269 ; VI-FLUSH-NEXT:    s_mov_b32 s10, -1
 270 ; VI-FLUSH-NEXT:    s_mov_b32 s14, s10
 271 ; VI-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 272 ; VI-FLUSH-NEXT:    s_mov_b32 s12, s6
 273 ; VI-FLUSH-NEXT:    s_mov_b32 s13, s7
 274 ; VI-FLUSH-NEXT:    s_mov_b32 s15, s11
 275 ; VI-FLUSH-NEXT:    s_mov_b32 s2, s10
 276 ; VI-FLUSH-NEXT:    s_mov_b32 s3, s11
 277 ; VI-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 278 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 279 ; VI-FLUSH-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc
 280 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 281 ; VI-FLUSH-NEXT:    s_mov_b32 s8, s4
 282 ; VI-FLUSH-NEXT:    s_mov_b32 s9, s5
 283 ; VI-FLUSH-NEXT:    v_madmk_f16 v0, v0, 0x4200, v1
 284 ; VI-FLUSH-NEXT:    buffer_store_short v0, off, s[8:11], 0
 285 ; VI-FLUSH-NEXT:    s_endpgm
 286 ;
 287 ; VI-DENORM-LABEL: fmuladd_f16_imm_a:
 288 ; VI-DENORM:       ; %bb.0:
 289 ; VI-DENORM-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 290 ; VI-DENORM-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 291 ; VI-DENORM-NEXT:    s_mov_b32 s11, 0xf000
 292 ; VI-DENORM-NEXT:    s_mov_b32 s10, -1
 293 ; VI-DENORM-NEXT:    s_mov_b32 s14, s10
 294 ; VI-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 295 ; VI-DENORM-NEXT:    s_mov_b32 s12, s6
 296 ; VI-DENORM-NEXT:    s_mov_b32 s13, s7
 297 ; VI-DENORM-NEXT:    s_mov_b32 s15, s11
 298 ; VI-DENORM-NEXT:    s_mov_b32 s2, s10
 299 ; VI-DENORM-NEXT:    s_mov_b32 s3, s11
 300 ; VI-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 301 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 302 ; VI-DENORM-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc
 303 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 304 ; VI-DENORM-NEXT:    s_movk_i32 s0, 0x4200
 305 ; VI-DENORM-NEXT:    s_mov_b32 s8, s4
 306 ; VI-DENORM-NEXT:    s_mov_b32 s9, s5
 307 ; VI-DENORM-NEXT:    v_fma_f16 v0, v0, s0, v1
 308 ; VI-DENORM-NEXT:    buffer_store_short v0, off, s[8:11], 0
 309 ; VI-DENORM-NEXT:    s_endpgm
 310 ;
 311 ; GFX10-FLUSH-LABEL: fmuladd_f16_imm_a:
 312 ; GFX10-FLUSH:       ; %bb.0:
 313 ; GFX10-FLUSH-NEXT:    s_clause 0x1
 314 ; GFX10-FLUSH-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 315 ; GFX10-FLUSH-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 316 ; GFX10-FLUSH-NEXT:    s_mov_b32 s10, -1
 317 ; GFX10-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 318 ; GFX10-FLUSH-NEXT:    s_mov_b32 s14, s10
 319 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, s11
 320 ; GFX10-FLUSH-NEXT:    s_mov_b32 s2, s10
 321 ; GFX10-FLUSH-NEXT:    s_mov_b32 s3, s11
 322 ; GFX10-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 323 ; GFX10-FLUSH-NEXT:    s_mov_b32 s12, s6
 324 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s7
 325 ; GFX10-FLUSH-NEXT:    s_mov_b32 s8, s4
 326 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 327 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 328 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc dlc
 329 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 330 ; GFX10-FLUSH-NEXT:    s_mov_b32 s9, s5
 331 ; GFX10-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 332 ; GFX10-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 333 ; GFX10-FLUSH-NEXT:    buffer_store_short v0, off, s[8:11], 0
 334 ; GFX10-FLUSH-NEXT:    s_endpgm
 335 ;
 336 ; GFX10-DENORM-LABEL: fmuladd_f16_imm_a:
 337 ; GFX10-DENORM:       ; %bb.0:
 338 ; GFX10-DENORM-NEXT:    s_clause 0x1
 339 ; GFX10-DENORM-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 340 ; GFX10-DENORM-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 341 ; GFX10-DENORM-NEXT:    s_mov_b32 s10, -1
 342 ; GFX10-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 343 ; GFX10-DENORM-NEXT:    s_mov_b32 s14, s10
 344 ; GFX10-DENORM-NEXT:    s_mov_b32 s15, s11
 345 ; GFX10-DENORM-NEXT:    s_mov_b32 s2, s10
 346 ; GFX10-DENORM-NEXT:    s_mov_b32 s3, s11
 347 ; GFX10-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 348 ; GFX10-DENORM-NEXT:    s_mov_b32 s12, s6
 349 ; GFX10-DENORM-NEXT:    s_mov_b32 s13, s7
 350 ; GFX10-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 351 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 352 ; GFX10-DENORM-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc dlc
 353 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 354 ; GFX10-DENORM-NEXT:    s_mov_b32 s8, s4
 355 ; GFX10-DENORM-NEXT:    s_mov_b32 s9, s5
 356 ; GFX10-DENORM-NEXT:    v_fmamk_f16 v0, v0, 0x4200, v1
 357 ; GFX10-DENORM-NEXT:    buffer_store_short v0, off, s[8:11], 0
 358 ; GFX10-DENORM-NEXT:    s_endpgm
 359 ;
 360 ; GFX11-FLUSH-LABEL: fmuladd_f16_imm_a:
 361 ; GFX11-FLUSH:       ; %bb.0:
 362 ; GFX11-FLUSH-NEXT:    s_clause 0x1
 363 ; GFX11-FLUSH-NEXT:    s_load_b128 s[4:7], s[2:3], 0x24
 364 ; GFX11-FLUSH-NEXT:    s_load_b64 s[0:1], s[2:3], 0x34
 365 ; GFX11-FLUSH-NEXT:    s_mov_b32 s10, -1
 366 ; GFX11-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 367 ; GFX11-FLUSH-NEXT:    s_mov_b32 s14, s10
 368 ; GFX11-FLUSH-NEXT:    s_mov_b32 s15, s11
 369 ; GFX11-FLUSH-NEXT:    s_mov_b32 s2, s10
 370 ; GFX11-FLUSH-NEXT:    s_mov_b32 s3, s11
 371 ; GFX11-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 372 ; GFX11-FLUSH-NEXT:    s_mov_b32 s12, s6
 373 ; GFX11-FLUSH-NEXT:    s_mov_b32 s13, s7
 374 ; GFX11-FLUSH-NEXT:    s_mov_b32 s8, s4
 375 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 376 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 377 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v1, off, s[0:3], 0 glc dlc
 378 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 379 ; GFX11-FLUSH-NEXT:    s_mov_b32 s9, s5
 380 ; GFX11-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 381 ; GFX11-FLUSH-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 382 ; GFX11-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 383 ; GFX11-FLUSH-NEXT:    buffer_store_b16 v0, off, s[8:11], 0
 384 ; GFX11-FLUSH-NEXT:    s_nop 0
 385 ; GFX11-FLUSH-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 386 ; GFX11-FLUSH-NEXT:    s_endpgm
 387 ;
 388 ; GFX11-DENORM-LABEL: fmuladd_f16_imm_a:
 389 ; GFX11-DENORM:       ; %bb.0:
 390 ; GFX11-DENORM-NEXT:    s_clause 0x1
 391 ; GFX11-DENORM-NEXT:    s_load_b128 s[4:7], s[2:3], 0x24
 392 ; GFX11-DENORM-NEXT:    s_load_b64 s[0:1], s[2:3], 0x34
 393 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 394 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 395 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 396 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 397 ; GFX11-DENORM-NEXT:    s_mov_b32 s2, s10
 398 ; GFX11-DENORM-NEXT:    s_mov_b32 s3, s11
 399 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 400 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s6
 401 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s7
 402 ; GFX11-DENORM-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 403 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 404 ; GFX11-DENORM-NEXT:    buffer_load_u16 v1, off, s[0:3], 0 glc dlc
 405 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 406 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s4
 407 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s5
 408 ; GFX11-DENORM-NEXT:    v_fmac_f16_e32 v1, 0x4200, v0
 409 ; GFX11-DENORM-NEXT:    buffer_store_b16 v1, off, s[8:11], 0
 410 ; GFX11-DENORM-NEXT:    s_nop 0
 411 ; GFX11-DENORM-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 412 ; GFX11-DENORM-NEXT:    s_endpgm
 413     ptr addrspace(1) %r,
 414     ptr addrspace(1) %b,
 415     ptr addrspace(1) %c) {
 416   %b.val = load volatile half, ptr addrspace(1) %b
 417   %c.val = load volatile half, ptr addrspace(1) %c
 418   %r.val = call half @llvm.fmuladd.f16(half 3.0, half %b.val, half %c.val)
 419   store half %r.val, ptr addrspace(1) %r
 420   ret void
 421 }
 422
 423 define amdgpu_kernel void @fmuladd_f16_imm_b(
 424 ; SI-LABEL: fmuladd_f16_imm_b:
 425 ; SI:       ; %bb.0:
 426 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
 427 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xd
 428 ; SI-NEXT:    s_mov_b32 s11, 0xf000
 429 ; SI-NEXT:    s_mov_b32 s10, -1
 430 ; SI-NEXT:    s_mov_b32 s14, s10
 431 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 432 ; SI-NEXT:    s_mov_b32 s12, s6
 433 ; SI-NEXT:    s_mov_b32 s13, s7
 434 ; SI-NEXT:    s_mov_b32 s15, s11
 435 ; SI-NEXT:    s_mov_b32 s2, s10
 436 ; SI-NEXT:    s_mov_b32 s3, s11
 437 ; SI-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 438 ; SI-NEXT:    s_waitcnt vmcnt(0)
 439 ; SI-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc
 440 ; SI-NEXT:    s_waitcnt vmcnt(0)
 441 ; SI-NEXT:    s_mov_b32 s8, s4
 442 ; SI-NEXT:    s_mov_b32 s9, s5
 443 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 444 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 445 ; SI-NEXT:    v_madmk_f32 v0, v0, 0x40400000, v1
 446 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 447 ; SI-NEXT:    buffer_store_short v0, off, s[8:11], 0
 448 ; SI-NEXT:    s_endpgm
 449 ;
 450 ; VI-FLUSH-LABEL: fmuladd_f16_imm_b:
 451 ; VI-FLUSH:       ; %bb.0:
 452 ; VI-FLUSH-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 453 ; VI-FLUSH-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 454 ; VI-FLUSH-NEXT:    s_mov_b32 s11, 0xf000
 455 ; VI-FLUSH-NEXT:    s_mov_b32 s10, -1
 456 ; VI-FLUSH-NEXT:    s_mov_b32 s14, s10
 457 ; VI-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 458 ; VI-FLUSH-NEXT:    s_mov_b32 s12, s6
 459 ; VI-FLUSH-NEXT:    s_mov_b32 s13, s7
 460 ; VI-FLUSH-NEXT:    s_mov_b32 s15, s11
 461 ; VI-FLUSH-NEXT:    s_mov_b32 s2, s10
 462 ; VI-FLUSH-NEXT:    s_mov_b32 s3, s11
 463 ; VI-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 464 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 465 ; VI-FLUSH-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc
 466 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 467 ; VI-FLUSH-NEXT:    s_mov_b32 s8, s4
 468 ; VI-FLUSH-NEXT:    s_mov_b32 s9, s5
 469 ; VI-FLUSH-NEXT:    v_madmk_f16 v0, v0, 0x4200, v1
 470 ; VI-FLUSH-NEXT:    buffer_store_short v0, off, s[8:11], 0
 471 ; VI-FLUSH-NEXT:    s_endpgm
 472 ;
 473 ; VI-DENORM-LABEL: fmuladd_f16_imm_b:
 474 ; VI-DENORM:       ; %bb.0:
 475 ; VI-DENORM-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 476 ; VI-DENORM-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 477 ; VI-DENORM-NEXT:    s_mov_b32 s11, 0xf000
 478 ; VI-DENORM-NEXT:    s_mov_b32 s10, -1
 479 ; VI-DENORM-NEXT:    s_mov_b32 s14, s10
 480 ; VI-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 481 ; VI-DENORM-NEXT:    s_mov_b32 s12, s6
 482 ; VI-DENORM-NEXT:    s_mov_b32 s13, s7
 483 ; VI-DENORM-NEXT:    s_mov_b32 s15, s11
 484 ; VI-DENORM-NEXT:    s_mov_b32 s2, s10
 485 ; VI-DENORM-NEXT:    s_mov_b32 s3, s11
 486 ; VI-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc
 487 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 488 ; VI-DENORM-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc
 489 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 490 ; VI-DENORM-NEXT:    s_movk_i32 s0, 0x4200
 491 ; VI-DENORM-NEXT:    s_mov_b32 s8, s4
 492 ; VI-DENORM-NEXT:    s_mov_b32 s9, s5
 493 ; VI-DENORM-NEXT:    v_fma_f16 v0, v0, s0, v1
 494 ; VI-DENORM-NEXT:    buffer_store_short v0, off, s[8:11], 0
 495 ; VI-DENORM-NEXT:    s_endpgm
 496 ;
 497 ; GFX10-FLUSH-LABEL: fmuladd_f16_imm_b:
 498 ; GFX10-FLUSH:       ; %bb.0:
 499 ; GFX10-FLUSH-NEXT:    s_clause 0x1
 500 ; GFX10-FLUSH-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 501 ; GFX10-FLUSH-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 502 ; GFX10-FLUSH-NEXT:    s_mov_b32 s10, -1
 503 ; GFX10-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 504 ; GFX10-FLUSH-NEXT:    s_mov_b32 s14, s10
 505 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, s11
 506 ; GFX10-FLUSH-NEXT:    s_mov_b32 s2, s10
 507 ; GFX10-FLUSH-NEXT:    s_mov_b32 s3, s11
 508 ; GFX10-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 509 ; GFX10-FLUSH-NEXT:    s_mov_b32 s12, s6
 510 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s7
 511 ; GFX10-FLUSH-NEXT:    s_mov_b32 s8, s4
 512 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 513 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 514 ; GFX10-FLUSH-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc dlc
 515 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 516 ; GFX10-FLUSH-NEXT:    s_mov_b32 s9, s5
 517 ; GFX10-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 518 ; GFX10-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 519 ; GFX10-FLUSH-NEXT:    buffer_store_short v0, off, s[8:11], 0
 520 ; GFX10-FLUSH-NEXT:    s_endpgm
 521 ;
 522 ; GFX10-DENORM-LABEL: fmuladd_f16_imm_b:
 523 ; GFX10-DENORM:       ; %bb.0:
 524 ; GFX10-DENORM-NEXT:    s_clause 0x1
 525 ; GFX10-DENORM-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 526 ; GFX10-DENORM-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
 527 ; GFX10-DENORM-NEXT:    s_mov_b32 s10, -1
 528 ; GFX10-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 529 ; GFX10-DENORM-NEXT:    s_mov_b32 s14, s10
 530 ; GFX10-DENORM-NEXT:    s_mov_b32 s15, s11
 531 ; GFX10-DENORM-NEXT:    s_mov_b32 s2, s10
 532 ; GFX10-DENORM-NEXT:    s_mov_b32 s3, s11
 533 ; GFX10-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 534 ; GFX10-DENORM-NEXT:    s_mov_b32 s12, s6
 535 ; GFX10-DENORM-NEXT:    s_mov_b32 s13, s7
 536 ; GFX10-DENORM-NEXT:    buffer_load_ushort v0, off, s[12:15], 0 glc dlc
 537 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 538 ; GFX10-DENORM-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 glc dlc
 539 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 540 ; GFX10-DENORM-NEXT:    s_mov_b32 s8, s4
 541 ; GFX10-DENORM-NEXT:    s_mov_b32 s9, s5
 542 ; GFX10-DENORM-NEXT:    v_fmamk_f16 v0, v0, 0x4200, v1
 543 ; GFX10-DENORM-NEXT:    buffer_store_short v0, off, s[8:11], 0
 544 ; GFX10-DENORM-NEXT:    s_endpgm
 545 ;
 546 ; GFX11-FLUSH-LABEL: fmuladd_f16_imm_b:
 547 ; GFX11-FLUSH:       ; %bb.0:
 548 ; GFX11-FLUSH-NEXT:    s_clause 0x1
 549 ; GFX11-FLUSH-NEXT:    s_load_b128 s[4:7], s[2:3], 0x24
 550 ; GFX11-FLUSH-NEXT:    s_load_b64 s[0:1], s[2:3], 0x34
 551 ; GFX11-FLUSH-NEXT:    s_mov_b32 s10, -1
 552 ; GFX11-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 553 ; GFX11-FLUSH-NEXT:    s_mov_b32 s14, s10
 554 ; GFX11-FLUSH-NEXT:    s_mov_b32 s15, s11
 555 ; GFX11-FLUSH-NEXT:    s_mov_b32 s2, s10
 556 ; GFX11-FLUSH-NEXT:    s_mov_b32 s3, s11
 557 ; GFX11-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 558 ; GFX11-FLUSH-NEXT:    s_mov_b32 s12, s6
 559 ; GFX11-FLUSH-NEXT:    s_mov_b32 s13, s7
 560 ; GFX11-FLUSH-NEXT:    s_mov_b32 s8, s4
 561 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 562 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 563 ; GFX11-FLUSH-NEXT:    buffer_load_u16 v1, off, s[0:3], 0 glc dlc
 564 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 565 ; GFX11-FLUSH-NEXT:    s_mov_b32 s9, s5
 566 ; GFX11-FLUSH-NEXT:    v_mul_f16_e32 v0, 0x4200, v0
 567 ; GFX11-FLUSH-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 568 ; GFX11-FLUSH-NEXT:    v_add_f16_e32 v0, v0, v1
 569 ; GFX11-FLUSH-NEXT:    buffer_store_b16 v0, off, s[8:11], 0
 570 ; GFX11-FLUSH-NEXT:    s_nop 0
 571 ; GFX11-FLUSH-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 572 ; GFX11-FLUSH-NEXT:    s_endpgm
 573 ;
 574 ; GFX11-DENORM-LABEL: fmuladd_f16_imm_b:
 575 ; GFX11-DENORM:       ; %bb.0:
 576 ; GFX11-DENORM-NEXT:    s_clause 0x1
 577 ; GFX11-DENORM-NEXT:    s_load_b128 s[4:7], s[2:3], 0x24
 578 ; GFX11-DENORM-NEXT:    s_load_b64 s[0:1], s[2:3], 0x34
 579 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 580 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 581 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 582 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 583 ; GFX11-DENORM-NEXT:    s_mov_b32 s2, s10
 584 ; GFX11-DENORM-NEXT:    s_mov_b32 s3, s11
 585 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 586 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s6
 587 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s7
 588 ; GFX11-DENORM-NEXT:    buffer_load_u16 v0, off, s[12:15], 0 glc dlc
 589 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 590 ; GFX11-DENORM-NEXT:    buffer_load_u16 v1, off, s[0:3], 0 glc dlc
 591 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 592 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s4
 593 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s5
 594 ; GFX11-DENORM-NEXT:    v_fmac_f16_e32 v1, 0x4200, v0
 595 ; GFX11-DENORM-NEXT:    buffer_store_b16 v1, off, s[8:11], 0
 596 ; GFX11-DENORM-NEXT:    s_nop 0
 597 ; GFX11-DENORM-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 598 ; GFX11-DENORM-NEXT:    s_endpgm
 599     ptr addrspace(1) %r,
 600     ptr addrspace(1) %a,
 601     ptr addrspace(1) %c) {
 602   %a.val = load volatile half, ptr addrspace(1) %a
 603   %c.val = load volatile half, ptr addrspace(1) %c
 604   %r.val = call half @llvm.fmuladd.f16(half %a.val, half 3.0, half %c.val)
 605   store half %r.val, ptr addrspace(1) %r
 606   ret void
 607 }
 608
 609 define amdgpu_kernel void @fmuladd_v2f16(
 610 ; SI-LABEL: fmuladd_v2f16:
 611 ; SI:       ; %bb.0:
 612 ; SI-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x9
 613 ; SI-NEXT:    s_mov_b32 s11, 0xf000
 614 ; SI-NEXT:    s_mov_b32 s10, -1
 615 ; SI-NEXT:    s_mov_b32 s14, s10
 616 ; SI-NEXT:    s_mov_b32 s15, s11
 617 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 618 ; SI-NEXT:    s_mov_b32 s12, s2
 619 ; SI-NEXT:    s_mov_b32 s13, s3
 620 ; SI-NEXT:    s_mov_b32 s16, s4
 621 ; SI-NEXT:    s_mov_b32 s17, s5
 622 ; SI-NEXT:    s_mov_b32 s18, s10
 623 ; SI-NEXT:    s_mov_b32 s19, s11
 624 ; SI-NEXT:    s_mov_b32 s4, s6
 625 ; SI-NEXT:    s_mov_b32 s5, s7
 626 ; SI-NEXT:    s_mov_b32 s6, s10
 627 ; SI-NEXT:    buffer_load_dword v0, off, s[12:15], 0
 628 ; SI-NEXT:    s_mov_b32 s7, s11
 629 ; SI-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 630 ; SI-NEXT:    buffer_load_dword v2, off, s[4:7], 0
 631 ; SI-NEXT:    s_mov_b32 s8, s0
 632 ; SI-NEXT:    s_mov_b32 s9, s1
 633 ; SI-NEXT:    s_waitcnt vmcnt(2)
 634 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v0
 635 ; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
 636 ; SI-NEXT:    s_waitcnt vmcnt(1)
 637 ; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
 638 ; SI-NEXT:    s_waitcnt vmcnt(0)
 639 ; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
 640 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 641 ; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
 642 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
 643 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 644 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 645 ; SI-NEXT:    v_mac_f32_e32 v5, v0, v4
 646 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v5
 647 ; SI-NEXT:    v_mac_f32_e32 v2, v3, v1
 648 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v2
 649 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 650 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 651 ; SI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
 652 ; SI-NEXT:    s_endpgm
 653 ;
 654 ; VI-FLUSH-LABEL: fmuladd_v2f16:
 655 ; VI-FLUSH:       ; %bb.0:
 656 ; VI-FLUSH-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x24
 657 ; VI-FLUSH-NEXT:    s_mov_b32 s11, 0xf000
 658 ; VI-FLUSH-NEXT:    s_mov_b32 s10, -1
 659 ; VI-FLUSH-NEXT:    s_mov_b32 s14, s10
 660 ; VI-FLUSH-NEXT:    s_mov_b32 s15, s11
 661 ; VI-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 662 ; VI-FLUSH-NEXT:    s_mov_b32 s12, s2
 663 ; VI-FLUSH-NEXT:    s_mov_b32 s13, s3
 664 ; VI-FLUSH-NEXT:    s_mov_b32 s16, s4
 665 ; VI-FLUSH-NEXT:    s_mov_b32 s17, s5
 666 ; VI-FLUSH-NEXT:    s_mov_b32 s18, s10
 667 ; VI-FLUSH-NEXT:    s_mov_b32 s19, s11
 668 ; VI-FLUSH-NEXT:    s_mov_b32 s4, s6
 669 ; VI-FLUSH-NEXT:    s_mov_b32 s5, s7
 670 ; VI-FLUSH-NEXT:    s_mov_b32 s6, s10
 671 ; VI-FLUSH-NEXT:    s_mov_b32 s7, s11
 672 ; VI-FLUSH-NEXT:    buffer_load_dword v0, off, s[12:15], 0
 673 ; VI-FLUSH-NEXT:    buffer_load_dword v1, off, s[4:7], 0
 674 ; VI-FLUSH-NEXT:    buffer_load_dword v2, off, s[16:19], 0
 675 ; VI-FLUSH-NEXT:    s_mov_b32 s8, s0
 676 ; VI-FLUSH-NEXT:    s_mov_b32 s9, s1
 677 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 678 ; VI-FLUSH-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
 679 ; VI-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 680 ; 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
 681 ; VI-FLUSH-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 682 ; VI-FLUSH-NEXT:    v_mac_f16_e32 v1, v0, v2
 683 ; VI-FLUSH-NEXT:    v_or_b32_e32 v0, v1, v3
 684 ; VI-FLUSH-NEXT:    buffer_store_dword v0, off, s[8:11], 0
 685 ; VI-FLUSH-NEXT:    s_endpgm
 686 ;
 687 ; VI-DENORM-LABEL: fmuladd_v2f16:
 688 ; VI-DENORM:       ; %bb.0:
 689 ; VI-DENORM-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x24
 690 ; VI-DENORM-NEXT:    s_mov_b32 s11, 0xf000
 691 ; VI-DENORM-NEXT:    s_mov_b32 s10, -1
 692 ; VI-DENORM-NEXT:    s_mov_b32 s14, s10
 693 ; VI-DENORM-NEXT:    s_mov_b32 s15, s11
 694 ; VI-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 695 ; VI-DENORM-NEXT:    s_mov_b32 s16, s4
 696 ; VI-DENORM-NEXT:    s_mov_b32 s17, s5
 697 ; VI-DENORM-NEXT:    s_mov_b32 s4, s6
 698 ; VI-DENORM-NEXT:    s_mov_b32 s5, s7
 699 ; VI-DENORM-NEXT:    s_mov_b32 s6, s10
 700 ; VI-DENORM-NEXT:    s_mov_b32 s7, s11
 701 ; VI-DENORM-NEXT:    s_mov_b32 s12, s2
 702 ; VI-DENORM-NEXT:    s_mov_b32 s13, s3
 703 ; VI-DENORM-NEXT:    s_mov_b32 s18, s10
 704 ; VI-DENORM-NEXT:    s_mov_b32 s19, s11
 705 ; VI-DENORM-NEXT:    buffer_load_dword v0, off, s[4:7], 0
 706 ; VI-DENORM-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 707 ; VI-DENORM-NEXT:    buffer_load_dword v2, off, s[12:15], 0
 708 ; VI-DENORM-NEXT:    s_mov_b32 s8, s0
 709 ; VI-DENORM-NEXT:    s_mov_b32 s9, s1
 710 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(2)
 711 ; VI-DENORM-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
 712 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(1)
 713 ; VI-DENORM-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
 714 ; VI-DENORM-NEXT:    s_waitcnt vmcnt(0)
 715 ; VI-DENORM-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
 716 ; VI-DENORM-NEXT:    v_fma_f16 v3, v5, v4, v3
 717 ; VI-DENORM-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 718 ; VI-DENORM-NEXT:    v_fma_f16 v0, v2, v1, v0
 719 ; VI-DENORM-NEXT:    v_or_b32_e32 v0, v0, v3
 720 ; VI-DENORM-NEXT:    buffer_store_dword v0, off, s[8:11], 0
 721 ; VI-DENORM-NEXT:    s_endpgm
 722 ;
 723 ; GFX10-FLUSH-LABEL: fmuladd_v2f16:
 724 ; GFX10-FLUSH:       ; %bb.0:
 725 ; GFX10-FLUSH-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
 726 ; GFX10-FLUSH-NEXT:    s_mov_b32 s2, -1
 727 ; GFX10-FLUSH-NEXT:    s_mov_b32 s3, 0x31016000
 728 ; GFX10-FLUSH-NEXT:    s_mov_b32 s14, s2
 729 ; GFX10-FLUSH-NEXT:    s_mov_b32 s15, s3
 730 ; GFX10-FLUSH-NEXT:    s_mov_b32 s18, s2
 731 ; GFX10-FLUSH-NEXT:    s_mov_b32 s19, s3
 732 ; GFX10-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 733 ; GFX10-FLUSH-NEXT:    s_mov_b32 s12, s6
 734 ; GFX10-FLUSH-NEXT:    s_mov_b32 s13, s7
 735 ; GFX10-FLUSH-NEXT:    s_mov_b32 s16, s8
 736 ; GFX10-FLUSH-NEXT:    s_mov_b32 s17, s9
 737 ; GFX10-FLUSH-NEXT:    buffer_load_dword v0, off, s[12:15], 0
 738 ; GFX10-FLUSH-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 739 ; GFX10-FLUSH-NEXT:    s_mov_b32 s8, s10
 740 ; GFX10-FLUSH-NEXT:    s_mov_b32 s9, s11
 741 ; GFX10-FLUSH-NEXT:    s_mov_b32 s10, s2
 742 ; GFX10-FLUSH-NEXT:    s_mov_b32 s11, s3
 743 ; GFX10-FLUSH-NEXT:    s_mov_b32 s0, s4
 744 ; GFX10-FLUSH-NEXT:    buffer_load_dword v2, off, s[8:11], 0
 745 ; GFX10-FLUSH-NEXT:    s_mov_b32 s1, s5
 746 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 747 ; GFX10-FLUSH-NEXT:    v_pk_mul_f16 v0, v0, v1
 748 ; GFX10-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 749 ; GFX10-FLUSH-NEXT:    v_pk_add_f16 v0, v0, v2
 750 ; GFX10-FLUSH-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 751 ; GFX10-FLUSH-NEXT:    s_endpgm
 752 ;
 753 ; GFX10-DENORM-LABEL: fmuladd_v2f16:
 754 ; GFX10-DENORM:       ; %bb.0:
 755 ; GFX10-DENORM-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
 756 ; GFX10-DENORM-NEXT:    s_mov_b32 s2, -1
 757 ; GFX10-DENORM-NEXT:    s_mov_b32 s3, 0x31016000
 758 ; GFX10-DENORM-NEXT:    s_mov_b32 s14, s2
 759 ; GFX10-DENORM-NEXT:    s_mov_b32 s15, s3
 760 ; GFX10-DENORM-NEXT:    s_mov_b32 s18, s2
 761 ; GFX10-DENORM-NEXT:    s_mov_b32 s19, s3
 762 ; GFX10-DENORM-NEXT:    s_mov_b32 s22, s2
 763 ; GFX10-DENORM-NEXT:    s_mov_b32 s23, s3
 764 ; GFX10-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 765 ; GFX10-DENORM-NEXT:    s_mov_b32 s12, s6
 766 ; GFX10-DENORM-NEXT:    s_mov_b32 s13, s7
 767 ; GFX10-DENORM-NEXT:    s_mov_b32 s16, s8
 768 ; GFX10-DENORM-NEXT:    s_mov_b32 s17, s9
 769 ; GFX10-DENORM-NEXT:    s_mov_b32 s20, s10
 770 ; GFX10-DENORM-NEXT:    s_mov_b32 s21, s11
 771 ; GFX10-DENORM-NEXT:    buffer_load_dword v0, off, s[12:15], 0
 772 ; GFX10-DENORM-NEXT:    buffer_load_dword v1, off, s[16:19], 0
 773 ; GFX10-DENORM-NEXT:    buffer_load_dword v2, off, s[20:23], 0
 774 ; GFX10-DENORM-NEXT:    s_mov_b32 s0, s4
 775 ; GFX10-DENORM-NEXT:    s_mov_b32 s1, s5
 776 ; GFX10-DENORM-NEXT:    s_waitcnt vmcnt(0)
 777 ; GFX10-DENORM-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
 778 ; GFX10-DENORM-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 779 ; GFX10-DENORM-NEXT:    s_endpgm
 780 ;
 781 ; GFX11-FLUSH-LABEL: fmuladd_v2f16:
 782 ; GFX11-FLUSH:       ; %bb.0:
 783 ; GFX11-FLUSH-NEXT:    s_load_b256 s[0:7], s[2:3], 0x24
 784 ; GFX11-FLUSH-NEXT:    s_mov_b32 s10, -1
 785 ; GFX11-FLUSH-NEXT:    s_mov_b32 s11, 0x31016000
 786 ; GFX11-FLUSH-NEXT:    s_mov_b32 s14, s10
 787 ; GFX11-FLUSH-NEXT:    s_mov_b32 s15, s11
 788 ; GFX11-FLUSH-NEXT:    s_mov_b32 s18, s10
 789 ; GFX11-FLUSH-NEXT:    s_mov_b32 s19, s11
 790 ; GFX11-FLUSH-NEXT:    s_waitcnt lgkmcnt(0)
 791 ; GFX11-FLUSH-NEXT:    s_mov_b32 s12, s2
 792 ; GFX11-FLUSH-NEXT:    s_mov_b32 s13, s3
 793 ; GFX11-FLUSH-NEXT:    s_mov_b32 s16, s4
 794 ; GFX11-FLUSH-NEXT:    s_mov_b32 s17, s5
 795 ; GFX11-FLUSH-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
 796 ; GFX11-FLUSH-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
 797 ; GFX11-FLUSH-NEXT:    s_mov_b32 s4, s6
 798 ; GFX11-FLUSH-NEXT:    s_mov_b32 s5, s7
 799 ; GFX11-FLUSH-NEXT:    s_mov_b32 s6, s10
 800 ; GFX11-FLUSH-NEXT:    s_mov_b32 s7, s11
 801 ; GFX11-FLUSH-NEXT:    s_mov_b32 s8, s0
 802 ; GFX11-FLUSH-NEXT:    buffer_load_b32 v2, off, s[4:7], 0
 803 ; GFX11-FLUSH-NEXT:    s_mov_b32 s9, s1
 804 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(1)
 805 ; GFX11-FLUSH-NEXT:    v_pk_mul_f16 v0, v0, v1
 806 ; GFX11-FLUSH-NEXT:    s_waitcnt vmcnt(0)
 807 ; GFX11-FLUSH-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 808 ; GFX11-FLUSH-NEXT:    v_pk_add_f16 v0, v0, v2
 809 ; GFX11-FLUSH-NEXT:    buffer_store_b32 v0, off, s[8:11], 0
 810 ; GFX11-FLUSH-NEXT:    s_nop 0
 811 ; GFX11-FLUSH-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 812 ; GFX11-FLUSH-NEXT:    s_endpgm
 813 ;
 814 ; GFX11-DENORM-LABEL: fmuladd_v2f16:
 815 ; GFX11-DENORM:       ; %bb.0:
 816 ; GFX11-DENORM-NEXT:    s_load_b256 s[0:7], s[2:3], 0x24
 817 ; GFX11-DENORM-NEXT:    s_mov_b32 s10, -1
 818 ; GFX11-DENORM-NEXT:    s_mov_b32 s11, 0x31016000
 819 ; GFX11-DENORM-NEXT:    s_mov_b32 s14, s10
 820 ; GFX11-DENORM-NEXT:    s_mov_b32 s15, s11
 821 ; GFX11-DENORM-NEXT:    s_mov_b32 s18, s10
 822 ; GFX11-DENORM-NEXT:    s_mov_b32 s19, s11
 823 ; GFX11-DENORM-NEXT:    s_mov_b32 s22, s10
 824 ; GFX11-DENORM-NEXT:    s_mov_b32 s23, s11
 825 ; GFX11-DENORM-NEXT:    s_waitcnt lgkmcnt(0)
 826 ; GFX11-DENORM-NEXT:    s_mov_b32 s12, s2
 827 ; GFX11-DENORM-NEXT:    s_mov_b32 s13, s3
 828 ; GFX11-DENORM-NEXT:    s_mov_b32 s16, s4
 829 ; GFX11-DENORM-NEXT:    s_mov_b32 s17, s5
 830 ; GFX11-DENORM-NEXT:    s_mov_b32 s20, s6
 831 ; GFX11-DENORM-NEXT:    s_mov_b32 s21, s7
 832 ; GFX11-DENORM-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
 833 ; GFX11-DENORM-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
 834 ; GFX11-DENORM-NEXT:    buffer_load_b32 v2, off, s[20:23], 0
 835 ; GFX11-DENORM-NEXT:    s_mov_b32 s8, s0
 836 ; GFX11-DENORM-NEXT:    s_mov_b32 s9, s1
 837 ; GFX11-DENORM-NEXT:    s_waitcnt vmcnt(0)
 838 ; GFX11-DENORM-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
 839 ; GFX11-DENORM-NEXT:    buffer_store_b32 v0, off, s[8:11], 0
 840 ; GFX11-DENORM-NEXT:    s_nop 0
 841 ; GFX11-DENORM-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 842 ; GFX11-DENORM-NEXT:    s_endpgm
 843     ptr addrspace(1) %r,
 844     ptr addrspace(1) %a,
 845     ptr addrspace(1) %b,
 846     ptr addrspace(1) %c) {
 847   %a.val = load <2 x half>, ptr addrspace(1) %a
 848   %b.val = load <2 x half>, ptr addrspace(1) %b
 849   %c.val = load <2 x half>, ptr addrspace(1) %c
 850   %r.val = call <2 x half> @llvm.fmuladd.v2f16(<2 x half> %a.val, <2 x half> %b.val, <2 x half> %c.val)
 851   store <2 x half> %r.val, ptr addrspace(1) %r
 852   ret void
 853 }