src/amd/compiler/aco_opcodes.py

   1 #
   2 # Copyright (c) 2018 Valve Corporation
   3 #
   4 # Permission is hereby granted, free of charge, to any person obtaining a
   5 # copy of this software and associated documentation files (the "Software"),
   6 # to deal in the Software without restriction, including without limitation
   7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 # and/or sell copies of the Software, and to permit persons to whom the
   9 # Software is furnished to do so, subject to the following conditions:
  10 #
  11 # The above copyright notice and this permission notice (including the next
  12 # paragraph) shall be included in all copies or substantial portions of the
  13 # Software.
  14 #
  15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 # IN THE SOFTWARE.
  22 #
  23 # Authors:
  24 #    Daniel Schuermann (daniel.schuermann@campus.tu-berlin.de)
  25
  26
  27 # Class that represents all the information we have about the opcode
  28 # NOTE: this must be kept in sync with aco_op_info
  29
  30 import sys
  31 from enum import Enum
  32
  33 class Format(Enum):
  34    PSEUDO = 0
  35    SOP1 = 1
  36    SOP2 = 2
  37    SOPK = 3
  38    SOPP = 4
  39    SOPC = 5
  40    SMEM = 6
  41    DS = 8
  42    MTBUF = 9
  43    MUBUF = 10
  44    MIMG = 11
  45    EXP = 12
  46    FLAT = 13
  47    GLOBAL = 14
  48    SCRATCH = 15
  49    PSEUDO_BRANCH = 16
  50    PSEUDO_BARRIER = 17
  51    PSEUDO_REDUCTION = 18
  52    VOP1 = 1 << 8
  53    VOP2 = 1 << 9
  54    VOPC = 1 << 10
  55    VOP3A = 1 << 11
  56    VOP3B = 1 << 11
  57    VOP3P = 1 << 12
  58    VINTRP = 1 << 13
  59    DPP = 1 << 14
  60    SDWA = 1 << 15
  61
  62    def get_builder_fields(self):
  63       if self == Format.SOPK:
  64          return [('uint16_t', 'imm', None)]
  65       elif self == Format.SOPP:
  66          return [('uint32_t', 'block', '-1'),
  67                  ('uint32_t', 'imm', '0')]
  68       elif self == Format.SMEM:
  69          return [('bool', 'can_reorder', 'true'),
  70                  ('bool', 'glc', 'false'),
  71                  ('bool', 'dlc', 'false'),
  72                  ('bool', 'nv', 'false')]
  73       elif self == Format.DS:
  74          return [('int16_t', 'offset0', '0'),
  75                  ('int8_t', 'offset1', '0'),
  76                  ('bool', 'gds', 'false')]
  77       elif self == Format.MTBUF:
  78          return [('unsigned', 'dfmt', None),
  79                  ('unsigned', 'nfmt', None),
  80                  ('unsigned', 'offset', None),
  81                  ('bool', 'offen', None),
  82                  ('bool', 'idxen', 'false'),
  83                  ('bool', 'disable_wqm', 'false'),
  84                  ('bool', 'glc', 'false'),
  85                  ('bool', 'dlc', 'false'),
  86                  ('bool', 'slc', 'false'),
  87                  ('bool', 'tfe', 'false')]
  88       elif self == Format.MUBUF:
  89          return [('unsigned', 'offset', None),
  90                  ('bool', 'offen', None),
  91                  ('bool', 'idxen', 'false'),
  92                  ('bool', 'addr64', 'false'),
  93                  ('bool', 'disable_wqm', 'false'),
  94                  ('bool', 'glc', 'false'),
  95                  ('bool', 'dlc', 'false'),
  96                  ('bool', 'slc', 'false'),
  97                  ('bool', 'tfe', 'false'),
  98                  ('bool', 'lds', 'false')]
  99       elif self == Format.MIMG:
 100          return [('unsigned', 'dmask', '0xF'),
 101                  ('bool', 'da', 'false'),
 102                  ('bool', 'unrm', 'true'),
 103                  ('bool', 'disable_wqm', 'false'),
 104                  ('bool', 'glc', 'false'),
 105                  ('bool', 'dlc', 'false'),
 106                  ('bool', 'slc', 'false'),
 107                  ('bool', 'tfe', 'false'),
 108                  ('bool', 'lwe', 'false'),
 109                  ('bool', 'r128_a16', 'false', 'r128'),
 110                  ('bool', 'd16', 'false')]
 111          return [('unsigned', 'attribute', None),
 112                  ('unsigned', 'component', None)]
 113       elif self == Format.EXP:
 114          return [('unsigned', 'enabled_mask', None),
 115                  ('unsigned', 'dest', None),
 116                  ('bool', 'compr', 'false', 'compressed'),
 117                  ('bool', 'done', 'false'),
 118                  ('bool', 'vm', 'false', 'valid_mask')]
 119       elif self == Format.PSEUDO_BRANCH:
 120          return [('uint32_t', 'target0', '0', 'target[0]'),
 121                  ('uint32_t', 'target1', '0', 'target[1]')]
 122       elif self == Format.PSEUDO_REDUCTION:
 123          return [('ReduceOp', 'op', None, 'reduce_op'),
 124                  ('unsigned', 'cluster_size', '0')]
 125       elif self == Format.VINTRP:
 126          return [('unsigned', 'attribute', None),
 127                  ('unsigned', 'component', None)]
 128       elif self == Format.DPP:
 129          return [('uint16_t', 'dpp_ctrl', None),
 130                  ('uint8_t', 'row_mask', '0xF'),
 131                  ('uint8_t', 'bank_mask', '0xF'),
 132                  ('bool', 'bound_ctrl', 'false')]
 133       elif self in [Format.FLAT, Format.GLOBAL, Format.SCRATCH]:
 134          return [('uint16_t', 'offset', 0),
 135                  ('bool', 'can_reorder', 'true'),
 136                  ('bool', 'glc', 'false'),
 137                  ('bool', 'slc', 'false'),
 138                  ('bool', 'lds', 'false'),
 139                  ('bool', 'nv', 'false')]
 140       else:
 141          return []
 142
 143    def get_builder_field_names(self):
 144       return [f[1] for f in self.get_builder_fields()]
 145
 146    def get_builder_field_dests(self):
 147       return [(f[3] if len(f) >= 4 else f[1]) for f in self.get_builder_fields()]
 148
 149    def get_builder_field_decls(self):
 150       return [('%s %s=%s' % (f[0], f[1], f[2]) if f[2] != None else '%s %s' % (f[0], f[1])) for f in self.get_builder_fields()]
 151
 152
 153 class Opcode(object):
 154    """Class that represents all the information we have about the opcode
 155    NOTE: this must be kept in sync with aco_op_info
 156    """
 157    def __init__(self, name, opcode_gfx7, opcode_gfx9, opcode_gfx10, format, input_mod, output_mod, is_atomic):
 158       """Parameters:
 159
 160       - name is the name of the opcode (prepend nir_op_ for the enum name)
 161       - all types are strings that get nir_type_ prepended to them
 162       - input_types is a list of types
 163       - algebraic_properties is a space-seperated string, where nir_op_is_ is
 164         prepended before each entry
 165       - const_expr is an expression or series of statements that computes the
 166         constant value of the opcode given the constant values of its inputs.
 167       """
 168       assert isinstance(name, str)
 169       assert isinstance(opcode_gfx7, int)
 170       assert isinstance(opcode_gfx9, int)
 171       assert isinstance(opcode_gfx10, int)
 172       assert isinstance(format, Format)
 173       assert isinstance(input_mod, bool)
 174       assert isinstance(output_mod, bool)
 175
 176       self.name = name
 177       self.opcode_gfx7 = opcode_gfx7
 178       self.opcode_gfx9 = opcode_gfx9
 179       self.opcode_gfx10 = opcode_gfx10
 180       self.input_mod = "1" if input_mod else "0"
 181       self.output_mod = "1" if output_mod else "0"
 182       self.is_atomic = "1" if is_atomic else "0"
 183       self.format = format
 184
 185
 186 # global dictionary of opcodes
 187 opcodes = {}
 188
 189 def opcode(name, opcode_gfx7 = -1, opcode_gfx9 = -1, opcode_gfx10 = -1, format = Format.PSEUDO, input_mod = False, output_mod = False, is_atomic = False):
 190    assert name not in opcodes
 191    opcodes[name] = Opcode(name, opcode_gfx7, opcode_gfx9, opcode_gfx10, format, input_mod, output_mod, is_atomic)
 192
 193 opcode("exp", 0, 0, 0, format = Format.EXP)
 194 opcode("p_parallelcopy")
 195 opcode("p_startpgm")
 196 opcode("p_phi")
 197 opcode("p_linear_phi")
 198 opcode("p_as_uniform")
 199
 200 opcode("p_create_vector")
 201 opcode("p_extract_vector")
 202 opcode("p_split_vector")
 203
 204 # start/end the parts where we can use exec based instructions
 205 # implicitly
 206 opcode("p_logical_start")
 207 opcode("p_logical_end")
 208
 209 # e.g. subgroupMin() in SPIR-V
 210 opcode("p_reduce", format=Format.PSEUDO_REDUCTION)
 211 # e.g. subgroupInclusiveMin()
 212 opcode("p_inclusive_scan", format=Format.PSEUDO_REDUCTION)
 213 # e.g. subgroupExclusiveMin()
 214 opcode("p_exclusive_scan", format=Format.PSEUDO_REDUCTION)
 215 # simulates proper bpermute behavior on GFX10 wave64
 216 opcode("p_wave64_bpermute", format=Format.PSEUDO_REDUCTION)
 217
 218 opcode("p_branch", format=Format.PSEUDO_BRANCH)
 219 opcode("p_cbranch", format=Format.PSEUDO_BRANCH)
 220 opcode("p_cbranch_z", format=Format.PSEUDO_BRANCH)
 221 opcode("p_cbranch_nz", format=Format.PSEUDO_BRANCH)
 222
 223 opcode("p_memory_barrier_common", format=Format.PSEUDO_BARRIER) # atomic, buffer, image and shared
 224 opcode("p_memory_barrier_atomic", format=Format.PSEUDO_BARRIER)
 225 opcode("p_memory_barrier_buffer", format=Format.PSEUDO_BARRIER)
 226 opcode("p_memory_barrier_image", format=Format.PSEUDO_BARRIER)
 227 opcode("p_memory_barrier_shared", format=Format.PSEUDO_BARRIER)
 228 opcode("p_memory_barrier_gs_data", format=Format.PSEUDO_BARRIER)
 229 opcode("p_memory_barrier_gs_sendmsg", format=Format.PSEUDO_BARRIER)
 230
 231 opcode("p_spill")
 232 opcode("p_reload")
 233
 234 # start/end linear vgprs
 235 opcode("p_start_linear_vgpr")
 236 opcode("p_end_linear_vgpr")
 237
 238 opcode("p_wqm")
 239 opcode("p_discard_if")
 240 opcode("p_load_helper")
 241 opcode("p_demote_to_helper")
 242 opcode("p_is_helper")
 243 opcode("p_exit_early_if")
 244
 245 opcode("p_fs_buffer_store_smem", format=Format.SMEM)
 246
 247
 248 # SOP2 instructions: 2 scalar inputs, 1 scalar output (+optional scc)
 249 SOP2 = {
 250   # GFX6, GFX7, GFX8, GFX9, GFX10, name
 251    (0x00, 0x00, 0x00, 0x00, 0x00, "s_add_u32"),
 252    (0x01, 0x01, 0x01, 0x01, 0x01, "s_sub_u32"),
 253    (0x02, 0x02, 0x02, 0x02, 0x02, "s_add_i32"),
 254    (0x03, 0x03, 0x03, 0x03, 0x03, "s_sub_i32"),
 255    (0x04, 0x04, 0x04, 0x04, 0x04, "s_addc_u32"),
 256    (0x05, 0x05, 0x05, 0x05, 0x05, "s_subb_u32"),
 257    (0x06, 0x06, 0x06, 0x06, 0x06, "s_min_i32"),
 258    (0x07, 0x07, 0x07, 0x07, 0x07, "s_min_u32"),
 259    (0x08, 0x08, 0x08, 0x08, 0x08, "s_max_i32"),
 260    (0x09, 0x09, 0x09, 0x09, 0x09, "s_max_u32"),
 261    (0x0a, 0x0a, 0x0a, 0x0a, 0x0a, "s_cselect_b32"),
 262    (0x0b, 0x0b, 0x0b, 0x0b, 0x0b, "s_cselect_b64"),
 263    (0x0e, 0x0e, 0x0c, 0x0c, 0x0e, "s_and_b32"),
 264    (0x0f, 0x0f, 0x0d, 0x0d, 0x0f, "s_and_b64"),
 265    (0x10, 0x10, 0x0e, 0x0e, 0x10, "s_or_b32"),
 266    (0x11, 0x11, 0x0f, 0x0f, 0x11, "s_or_b64"),
 267    (0x12, 0x12, 0x10, 0x10, 0x12, "s_xor_b32"),
 268    (0x13, 0x13, 0x11, 0x11, 0x13, "s_xor_b64"),
 269    (0x14, 0x14, 0x12, 0x12, 0x14, "s_andn2_b32"),
 270    (0x15, 0x15, 0x13, 0x13, 0x15, "s_andn2_b64"),
 271    (0x16, 0x16, 0x14, 0x14, 0x16, "s_orn2_b32"),
 272    (0x17, 0x17, 0x15, 0x15, 0x17, "s_orn2_b64"),
 273    (0x18, 0x18, 0x16, 0x16, 0x18, "s_nand_b32"),
 274    (0x19, 0x19, 0x17, 0x17, 0x19, "s_nand_b64"),
 275    (0x1a, 0x1a, 0x18, 0x18, 0x1a, "s_nor_b32"),
 276    (0x1b, 0x1b, 0x19, 0x19, 0x1b, "s_nor_b64"),
 277    (0x1c, 0x1c, 0x1a, 0x1a, 0x1c, "s_xnor_b32"),
 278    (0x1d, 0x1d, 0x1b, 0x1b, 0x1d, "s_xnor_b64"),
 279    (0x1e, 0x1e, 0x1c, 0x1c, 0x1e, "s_lshl_b32"),
 280    (0x1f, 0x1f, 0x1d, 0x1d, 0x1f, "s_lshl_b64"),
 281    (0x20, 0x20, 0x1e, 0x1e, 0x20, "s_lshr_b32"),
 282    (0x21, 0x21, 0x1f, 0x1f, 0x21, "s_lshr_b64"),
 283    (0x22, 0x22, 0x20, 0x20, 0x22, "s_ashr_i32"),
 284    (0x23, 0x23, 0x21, 0x21, 0x23, "s_ashr_i64"),
 285    (0x24, 0x24, 0x22, 0x22, 0x24, "s_bfm_b32"),
 286    (0x25, 0x25, 0x23, 0x23, 0x25, "s_bfm_b64"),
 287    (0x26, 0x26, 0x24, 0x24, 0x26, "s_mul_i32"),
 288    (0x27, 0x27, 0x25, 0x25, 0x27, "s_bfe_u32"),
 289    (0x28, 0x28, 0x26, 0x26, 0x28, "s_bfe_i32"),
 290    (0x29, 0x29, 0x27, 0x27, 0x29, "s_bfe_u64"),
 291    (0x2a, 0x2a, 0x28, 0x28, 0x2a, "s_bfe_i64"),
 292    (0x2b, 0x2b, 0x29, 0x29,   -1, "s_cbranch_g_fork"),
 293    (0x2c, 0x2c, 0x2a, 0x2a, 0x2c, "s_absdiff_i32"),
 294    (  -1,   -1, 0x2b, 0x2b,   -1, "s_rfe_restore_b64"),
 295    (  -1,   -1,   -1, 0x2e, 0x2e, "s_lshl1_add_u32"),
 296    (  -1,   -1,   -1, 0x2f, 0x2f, "s_lshl2_add_u32"),
 297    (  -1,   -1,   -1, 0x30, 0x30, "s_lshl3_add_u32"),
 298    (  -1,   -1,   -1, 0x31, 0x31, "s_lshl4_add_u32"),
 299    (  -1,   -1,   -1, 0x32, 0x32, "s_pack_ll_b32_b16"),
 300    (  -1,   -1,   -1, 0x33, 0x33, "s_pack_lh_b32_b16"),
 301    (  -1,   -1,   -1, 0x34, 0x34, "s_pack_hh_b32_b16"),
 302    (  -1,   -1,   -1, 0x2c, 0x35, "s_mul_hi_u32"),
 303    (  -1,   -1,   -1, 0x2d, 0x36, "s_mul_hi_i32"),
 304 }
 305 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in SOP2:
 306     opcode(name, gfx7, gfx9, gfx10, Format.SOP2)
 307
 308
 309 # SOPK instructions: 0 input (+ imm), 1 output + optional scc
 310 SOPK = {
 311   # GFX6, GFX7, GFX8, GFX9, GFX10, name
 312    (0x00, 0x00, 0x00, 0x00, 0x00, "s_movk_i32"),
 313    (  -1,   -1,   -1,   -1, 0x01, "s_version"), # GFX10+
 314    (0x02, 0x02, 0x01, 0x01, 0x02, "s_cmovk_i32"), # GFX8_GFX9
 315    (0x03, 0x03, 0x02, 0x02, 0x03, "s_cmpk_eq_i32"),
 316    (0x04, 0x04, 0x03, 0x03, 0x04, "s_cmpk_lg_i32"),
 317    (0x05, 0x05, 0x04, 0x04, 0x05, "s_cmpk_gt_i32"),
 318    (0x06, 0x06, 0x05, 0x05, 0x06, "s_cmpk_ge_i32"),
 319    (0x07, 0x07, 0x06, 0x06, 0x07, "s_cmpk_lt_i32"),
 320    (0x08, 0x08, 0x07, 0x07, 0x08, "s_cmpk_le_i32"),
 321    (0x09, 0x09, 0x08, 0x08, 0x09, "s_cmpk_eq_u32"),
 322    (0x0a, 0x0a, 0x09, 0x09, 0x0a, "s_cmpk_lg_u32"),
 323    (0x0b, 0x0b, 0x0a, 0x0a, 0x0b, "s_cmpk_gt_u32"),
 324    (0x0c, 0x0c, 0x0b, 0x0b, 0x0c, "s_cmpk_ge_u32"),
 325    (0x0d, 0x0d, 0x0c, 0x0c, 0x0d, "s_cmpk_lt_u32"),
 326    (0x0e, 0x0e, 0x0d, 0x0d, 0x0e, "s_cmpk_le_u32"),
 327    (0x0f, 0x0f, 0x0e, 0x0e, 0x0f, "s_addk_i32"),
 328    (0x10, 0x10, 0x0f, 0x0f, 0x10, "s_mulk_i32"),
 329    (0x11, 0x11, 0x10, 0x10,   -1, "s_cbranch_i_fork"),
 330    (0x12, 0x12, 0x11, 0x11, 0x12, "s_getreg_b32"),
 331    (0x13, 0x13, 0x12, 0x12, 0x13, "s_setreg_b32"),
 332    (0x15, 0x15, 0x14, 0x14, 0x15, "s_setreg_imm32_b32"), # requires 32bit literal
 333    (  -1,   -1, 0x15, 0x15, 0x16, "s_call_b64"),
 334    (  -1,   -1,   -1,   -1, 0x17, "s_waitcnt_vscnt"),
 335    (  -1,   -1,   -1,   -1, 0x18, "s_waitcnt_vmcnt"),
 336    (  -1,   -1,   -1,   -1, 0x19, "s_waitcnt_expcnt"),
 337    (  -1,   -1,   -1,   -1, 0x1a, "s_waitcnt_lgkmcnt"),
 338    (  -1,   -1,   -1,   -1, 0x1b, "s_subvector_loop_begin"),
 339    (  -1,   -1,   -1,   -1, 0x1c, "s_subvector_loop_end"),
 340 }
 341 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in SOPK:
 342    opcode(name, gfx7, gfx9, gfx10, Format.SOPK)
 343
 344
 345 # SOP1 instructions: 1 input, 1 output (+optional SCC)
 346 SOP1 = {
 347   # GFX6, GFX7, GFX8, GFX9, GFX10, name
 348    (0x03, 0x03, 0x00, 0x00, 0x03, "s_mov_b32"),
 349    (0x04, 0x04, 0x01, 0x01, 0x04, "s_mov_b64"),
 350    (0x05, 0x05, 0x02, 0x02, 0x05, "s_cmov_b32"),
 351    (0x06, 0x06, 0x03, 0x03, 0x06, "s_cmov_b64"),
 352    (0x07, 0x07, 0x04, 0x04, 0x07, "s_not_b32"),
 353    (0x08, 0x08, 0x05, 0x05, 0x08, "s_not_b64"),
 354    (0x09, 0x09, 0x06, 0x06, 0x09, "s_wqm_b32"),
 355    (0x0a, 0x0a, 0x07, 0x07, 0x0a, "s_wqm_b64"),
 356    (0x0b, 0x0b, 0x08, 0x08, 0x0b, "s_brev_b32"),
 357    (0x0c, 0x0c, 0x09, 0x09, 0x0c, "s_brev_b64"),
 358    (0x0d, 0x0d, 0x0a, 0x0a, 0x0d, "s_bcnt0_i32_b32"),
 359    (0x0e, 0x0e, 0x0b, 0x0b, 0x0e, "s_bcnt0_i32_b64"),
 360    (0x0f, 0x0f, 0x0c, 0x0c, 0x0f, "s_bcnt1_i32_b32"),
 361    (0x10, 0x10, 0x0d, 0x0d, 0x10, "s_bcnt1_i32_b64"),
 362    (0x11, 0x11, 0x0e, 0x0e, 0x11, "s_ff0_i32_b32"),
 363    (0x12, 0x12, 0x0f, 0x0f, 0x12, "s_ff0_i32_b64"),
 364    (0x13, 0x13, 0x10, 0x10, 0x13, "s_ff1_i32_b32"),
 365    (0x14, 0x14, 0x11, 0x11, 0x14, "s_ff1_i32_b64"),
 366    (0x15, 0x15, 0x12, 0x12, 0x15, "s_flbit_i32_b32"),
 367    (0x16, 0x16, 0x13, 0x13, 0x16, "s_flbit_i32_b64"),
 368    (0x17, 0x17, 0x14, 0x14, 0x17, "s_flbit_i32"),
 369    (0x18, 0x18, 0x15, 0x15, 0x18, "s_flbit_i32_i64"),
 370    (0x19, 0x19, 0x16, 0x16, 0x19, "s_sext_i32_i8"),
 371    (0x1a, 0x1a, 0x17, 0x17, 0x1a, "s_sext_i32_i16"),
 372    (0x1b, 0x1b, 0x18, 0x18, 0x1b, "s_bitset0_b32"),
 373    (0x1c, 0x1c, 0x19, 0x19, 0x1c, "s_bitset0_b64"),
 374    (0x1d, 0x1d, 0x1a, 0x1a, 0x1d, "s_bitset1_b32"),
 375    (0x1e, 0x1e, 0x1b, 0x1b, 0x1e, "s_bitset1_b64"),
 376    (0x1f, 0x1f, 0x1c, 0x1c, 0x1f, "s_getpc_b64"),
 377    (0x20, 0x20, 0x1d, 0x1d, 0x20, "s_setpc_b64"),
 378    (0x21, 0x21, 0x1e, 0x1e, 0x21, "s_swappc_b64"),
 379    (0x22, 0x22, 0x1f, 0x1f, 0x22, "s_rfe_b64"),
 380    (0x24, 0x24, 0x20, 0x20, 0x24, "s_and_saveexec_b64"),
 381    (0x25, 0x25, 0x21, 0x21, 0x25, "s_or_saveexec_b64"),
 382    (0x26, 0x26, 0x22, 0x22, 0x26, "s_xor_saveexec_b64"),
 383    (0x27, 0x27, 0x23, 0x23, 0x27, "s_andn2_saveexec_b64"),
 384    (0x28, 0x28, 0x24, 0x24, 0x28, "s_orn2_saveexec_b64"),
 385    (0x29, 0x29, 0x25, 0x25, 0x29, "s_nand_saveexec_b64"),
 386    (0x2a, 0x2a, 0x26, 0x26, 0x2a, "s_nor_saveexec_b64"),
 387    (0x2b, 0x2b, 0x27, 0x27, 0x2b, "s_xnor_saveexec_b64"),
 388    (0x2c, 0x2c, 0x28, 0x28, 0x2c, "s_quadmask_b32"),
 389    (0x2d, 0x2d, 0x29, 0x29, 0x2d, "s_quadmask_b64"),
 390    (0x2e, 0x2e, 0x2a, 0x2a, 0x2e, "s_movrels_b32"),
 391    (0x2f, 0x2f, 0x2b, 0x2b, 0x2f, "s_movrels_b64"),
 392    (0x30, 0x30, 0x2c, 0x2c, 0x30, "s_movreld_b32"),
 393    (0x31, 0x31, 0x2d, 0x2d, 0x31, "s_movreld_b64"),
 394    (0x32, 0x32, 0x2e, 0x2e,   -1, "s_cbranch_join"),
 395    (0x34, 0x34, 0x30, 0x30, 0x34, "s_abs_i32"),
 396    (0x35, 0x35,   -1,   -1, 0x35, "s_mov_fed_b32"),
 397    (  -1,   -1, 0x32, 0x32,   -1, "s_set_gpr_idx_idx"),
 398    (  -1,   -1,   -1, 0x33, 0x37, "s_andn1_saveexec_b64"),
 399    (  -1,   -1,   -1, 0x34, 0x38, "s_orn1_saveexec_b64"),
 400    (  -1,   -1,   -1, 0x35, 0x39, "s_andn1_wrexec_b64"),
 401    (  -1,   -1,   -1, 0x36, 0x3a, "s_andn2_wrexec_b64"),
 402    (  -1,   -1,   -1, 0x37, 0x3b, "s_bitreplicate_b64_b32"),
 403    (  -1,   -1,   -1,   -1, 0x3c, "s_and_saveexec_b32"),
 404    (  -1,   -1,   -1,   -1, 0x3d, "s_or_saveexec_b32"),
 405    (  -1,   -1,   -1,   -1, 0x3e, "s_xor_saveexec_b32"),
 406    (  -1,   -1,   -1,   -1, 0x3f, "s_andn2_saveexec_b32"),
 407    (  -1,   -1,   -1,   -1, 0x40, "s_orn2_saveexec_b32"),
 408    (  -1,   -1,   -1,   -1, 0x41, "s_nand_saveexec_b32"),
 409    (  -1,   -1,   -1,   -1, 0x42, "s_nor_saveexec_b32"),
 410    (  -1,   -1,   -1,   -1, 0x43, "s_xnor_saveexec_b32"),
 411    (  -1,   -1,   -1,   -1, 0x44, "s_andn1_saveexec_b32"),
 412    (  -1,   -1,   -1,   -1, 0x45, "s_orn1_saveexec_b32"),
 413    (  -1,   -1,   -1,   -1, 0x46, "s_andn1_wrexec_b32"),
 414    (  -1,   -1,   -1,   -1, 0x47, "s_andn2_wrexec_b32"),
 415    (  -1,   -1,   -1,   -1, 0x49, "s_movrelsd_2_b32"),
 416    # actually a pseudo-instruction. it's lowered to SALU during assembly though, so it's useful to identify it as a SOP1.
 417    (  -1,   -1,   -1,   -1,   -1, "p_constaddr"),
 418 }
 419 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in SOP1:
 420    opcode(name, gfx7, gfx9, gfx10, Format.SOP1)
 421
 422
 423 # SOPC instructions: 2 inputs and 0 outputs (+SCC)
 424 SOPC = {
 425   # GFX6, GFX7, GFX8, GFX9, GFX10, name
 426    (0x00, 0x00, 0x00, 0x00, 0x00, "s_cmp_eq_i32"),
 427    (0x01, 0x01, 0x01, 0x01, 0x01, "s_cmp_lg_i32"),
 428    (0x02, 0x02, 0x02, 0x02, 0x02, "s_cmp_gt_i32"),
 429    (0x03, 0x03, 0x03, 0x03, 0x03, "s_cmp_ge_i32"),
 430    (0x04, 0x04, 0x04, 0x04, 0x04, "s_cmp_lt_i32"),
 431    (0x05, 0x05, 0x05, 0x05, 0x05, "s_cmp_le_i32"),
 432    (0x06, 0x06, 0x06, 0x06, 0x06, "s_cmp_eq_u32"),
 433    (0x07, 0x07, 0x07, 0x07, 0x07, "s_cmp_lg_u32"),
 434    (0x08, 0x08, 0x08, 0x08, 0x08, "s_cmp_gt_u32"),
 435    (0x09, 0x09, 0x09, 0x09, 0x09, "s_cmp_ge_u32"),
 436    (0x0a, 0x0a, 0x0a, 0x0a, 0x0a, "s_cmp_lt_u32"),
 437    (0x0b, 0x0b, 0x0b, 0x0b, 0x0b, "s_cmp_le_u32"),
 438    (0x0c, 0x0c, 0x0c, 0x0c, 0x0c, "s_bitcmp0_b32"),
 439    (0x0d, 0x0d, 0x0d, 0x0d, 0x0d, "s_bitcmp1_b32"),
 440    (0x0e, 0x0e, 0x0e, 0x0e, 0x0e, "s_bitcmp0_b64"),
 441    (0x0f, 0x0f, 0x0f, 0x0f, 0x0f, "s_bitcmp1_b64"),
 442    (0x10, 0x10, 0x10, 0x10,   -1, "s_setvskip"),
 443    (  -1,   -1, 0x11, 0x11,   -1, "s_set_gpr_idx_on"),
 444    (  -1,   -1, 0x12, 0x12, 0x12, "s_cmp_eq_u64"),
 445    (  -1,   -1, 0x13, 0x13, 0x13, "s_cmp_lg_u64"),
 446 }
 447 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in SOPC:
 448    opcode(name, gfx7, gfx9, gfx10, Format.SOPC)
 449
 450
 451 # SOPP instructions: 0 inputs (+optional scc/vcc), 0 outputs
 452 SOPP = {
 453   # GFX6, GFX7, GFX8, GFX9, GFX10, name
 454    (0x00, 0x00, 0x00, 0x00, 0x00, "s_nop"),
 455    (0x01, 0x01, 0x01, 0x01, 0x01, "s_endpgm"),
 456    (0x02, 0x02, 0x02, 0x02, 0x02, "s_branch"),
 457    (  -1,   -1, 0x03, 0x03, 0x03, "s_wakeup"),
 458    (0x04, 0x04, 0x04, 0x04, 0x04, "s_cbranch_scc0"),
 459    (0x05, 0x05, 0x05, 0x05, 0x05, "s_cbranch_scc1"),
 460    (0x06, 0x06, 0x06, 0x06, 0x06, "s_cbranch_vccz"),
 461    (0x07, 0x07, 0x07, 0x07, 0x07, "s_cbranch_vccnz"),
 462    (0x08, 0x08, 0x08, 0x08, 0x08, "s_cbranch_execz"),
 463    (0x09, 0x09, 0x09, 0x09, 0x09, "s_cbranch_execnz"),
 464    (0x0a, 0x0a, 0x0a, 0x0a, 0x0a, "s_barrier"),
 465    (  -1, 0x0b, 0x0b, 0x0b, 0x0b, "s_setkill"),
 466    (0x0c, 0x0c, 0x0c, 0x0c, 0x0c, "s_waitcnt"),
 467    (0x0d, 0x0d, 0x0d, 0x0d, 0x0d, "s_sethalt"),
 468    (0x0e, 0x0e, 0x0e, 0x0e, 0x0e, "s_sleep"),
 469    (0x0f, 0x0f, 0x0f, 0x0f, 0x0f, "s_setprio"),
 470    (0x10, 0x10, 0x10, 0x10, 0x10, "s_sendmsg"),
 471    (0x11, 0x11, 0x11, 0x11, 0x11, "s_sendmsghalt"),
 472    (0x12, 0x12, 0x12, 0x12, 0x12, "s_trap"),
 473    (0x13, 0x13, 0x13, 0x13, 0x13, "s_icache_inv"),
 474    (0x14, 0x14, 0x14, 0x14, 0x14, "s_incperflevel"),
 475    (0x15, 0x15, 0x15, 0x15, 0x15, "s_decperflevel"),
 476    (0x16, 0x16, 0x16, 0x16, 0x16, "s_ttracedata"),
 477    (  -1, 0x17, 0x17, 0x17, 0x17, "s_cbranch_cdbgsys"),
 478    (  -1, 0x18, 0x18, 0x18, 0x18, "s_cbranch_cdbguser"),
 479    (  -1, 0x19, 0x19, 0x19, 0x19, "s_cbranch_cdbgsys_or_user"),
 480    (  -1, 0x1a, 0x1a, 0x1a, 0x1a, "s_cbranch_cdbgsys_and_user"),
 481    (  -1,   -1, 0x1b, 0x1b, 0x1b, "s_endpgm_saved"),
 482    (  -1,   -1, 0x1c, 0x1c,   -1, "s_set_gpr_idx_off"),
 483    (  -1,   -1, 0x1d, 0x1d,   -1, "s_set_gpr_idx_mode"),
 484    (  -1,   -1,   -1, 0x1e, 0x1e, "s_endpgm_ordered_ps_done"),
 485    (  -1,   -1,   -1,   -1, 0x1f, "s_code_end"),
 486    (  -1,   -1,   -1,   -1, 0x20, "s_inst_prefetch"),
 487    (  -1,   -1,   -1,   -1, 0x21, "s_clause"),
 488    (  -1,   -1,   -1,   -1, 0x22, "s_wait_idle"),
 489    (  -1,   -1,   -1,   -1, 0x23, "s_waitcnt_depctr"),
 490    (  -1,   -1,   -1,   -1, 0x24, "s_round_mode"),
 491    (  -1,   -1,   -1,   -1, 0x25, "s_denorm_mode"),
 492    (  -1,   -1,   -1,   -1, 0x26, "s_ttracedata_imm"),
 493 }
 494 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in SOPP:
 495    opcode(name, gfx7, gfx9, gfx10, Format.SOPP)
 496
 497
 498 # SMEM instructions: sbase input (2 sgpr), potentially 2 offset inputs, 1 sdata input/output
 499 SMEM = {
 500   # GFX6, GFX7, GFX8, GFX9, GFX10, name
 501    (0x00, 0x00, 0x00, 0x00, 0x00, "s_load_dword"),
 502    (0x01, 0x01, 0x01, 0x01, 0x01, "s_load_dwordx2"),
 503    (0x02, 0x02, 0x02, 0x02, 0x02, "s_load_dwordx4"),
 504    (0x03, 0x03, 0x03, 0x03, 0x03, "s_load_dwordx8"),
 505    (0x04, 0x04, 0x04, 0x04, 0x04, "s_load_dwordx16"),
 506    (  -1,   -1,   -1, 0x05, 0x05, "s_scratch_load_dword"),
 507    (  -1,   -1,   -1, 0x06, 0x06, "s_scratch_load_dwordx2"),
 508    (  -1,   -1,   -1, 0x07, 0x07, "s_scratch_load_dwordx4"),
 509    (0x08, 0x08, 0x08, 0x08, 0x08, "s_buffer_load_dword"),
 510    (0x09, 0x09, 0x09, 0x09, 0x09, "s_buffer_load_dwordx2"),
 511    (0x0a, 0x0a, 0x0a, 0x0a, 0x0a, "s_buffer_load_dwordx4"),
 512    (0x0b, 0x0b, 0x0b, 0x0b, 0x0b, "s_buffer_load_dwordx8"),
 513    (0x0c, 0x0c, 0x0c, 0x0c, 0x0c, "s_buffer_load_dwordx16"),
 514    (  -1,   -1, 0x10, 0x10, 0x10, "s_store_dword"),
 515    (  -1,   -1, 0x11, 0x11, 0x11, "s_store_dwordx2"),
 516    (  -1,   -1, 0x12, 0x12, 0x12, "s_store_dwordx4"),
 517    (  -1,   -1,   -1, 0x15, 0x15, "s_scratch_store_dword"),
 518    (  -1,   -1,   -1, 0x16, 0x16, "s_scratch_store_dwordx2"),
 519    (  -1,   -1,   -1, 0x17, 0x17, "s_scratch_store_dwordx4"),
 520    (  -1,   -1, 0x18, 0x18, 0x18, "s_buffer_store_dword"),
 521    (  -1,   -1, 0x19, 0x19, 0x19, "s_buffer_store_dwordx2"),
 522    (  -1,   -1, 0x1a, 0x1a, 0x1a, "s_buffer_store_dwordx4"),
 523    (  -1,   -1, 0x1f, 0x1f, 0x1f, "s_gl1_inv"),
 524    (0x1f, 0x1f, 0x20, 0x20, 0x20, "s_dcache_inv"),
 525    (  -1,   -1, 0x21, 0x21, 0x21, "s_dcache_wb"),
 526    (  -1, 0x1d, 0x22, 0x22,   -1, "s_dcache_inv_vol"),
 527    (  -1,   -1, 0x23, 0x23,   -1, "s_dcache_wb_vol"),
 528    (0x1e, 0x1e, 0x24, 0x24, 0x24, "s_memtime"),
 529    (  -1,   -1, 0x25, 0x25, 0x25, "s_memrealtime"),
 530    (  -1,   -1, 0x26, 0x26, 0x26, "s_atc_probe"),
 531    (  -1,   -1, 0x27, 0x27, 0x27, "s_atc_probe_buffer"),
 532    (  -1,   -1,   -1, 0x28, 0x28, "s_dcache_discard"),
 533    (  -1,   -1,   -1, 0x29, 0x29, "s_dcache_discard_x2"),
 534    (  -1,   -1,   -1,   -1, 0x2a, "s_get_waveid_in_workgroup"),
 535    (  -1,   -1,   -1, 0x40, 0x40, "s_buffer_atomic_swap"),
 536    (  -1,   -1,   -1, 0x41, 0x41, "s_buffer_atomic_cmpswap"),
 537    (  -1,   -1,   -1, 0x42, 0x42, "s_buffer_atomic_add"),
 538    (  -1,   -1,   -1, 0x43, 0x43, "s_buffer_atomic_sub"),
 539    (  -1,   -1,   -1, 0x44, 0x44, "s_buffer_atomic_smin"),
 540    (  -1,   -1,   -1, 0x45, 0x45, "s_buffer_atomic_umin"),
 541    (  -1,   -1,   -1, 0x46, 0x46, "s_buffer_atomic_smax"),
 542    (  -1,   -1,   -1, 0x47, 0x47, "s_buffer_atomic_umax"),
 543    (  -1,   -1,   -1, 0x48, 0x48, "s_buffer_atomic_and"),
 544    (  -1,   -1,   -1, 0x49, 0x49, "s_buffer_atomic_or"),
 545    (  -1,   -1,   -1, 0x4a, 0x4a, "s_buffer_atomic_xor"),
 546    (  -1,   -1,   -1, 0x4b, 0x4b, "s_buffer_atomic_inc"),
 547    (  -1,   -1,   -1, 0x4c, 0x4c, "s_buffer_atomic_dec"),
 548    (  -1,   -1,   -1, 0x60, 0x60, "s_buffer_atomic_swap_x2"),
 549    (  -1,   -1,   -1, 0x61, 0x61, "s_buffer_atomic_cmpswap_x2"),
 550    (  -1,   -1,   -1, 0x62, 0x62, "s_buffer_atomic_add_x2"),
 551    (  -1,   -1,   -1, 0x63, 0x63, "s_buffer_atomic_sub_x2"),
 552    (  -1,   -1,   -1, 0x64, 0x64, "s_buffer_atomic_smin_x2"),
 553    (  -1,   -1,   -1, 0x65, 0x65, "s_buffer_atomic_umin_x2"),
 554    (  -1,   -1,   -1, 0x66, 0x66, "s_buffer_atomic_smax_x2"),
 555    (  -1,   -1,   -1, 0x67, 0x67, "s_buffer_atomic_umax_x2"),
 556    (  -1,   -1,   -1, 0x68, 0x68, "s_buffer_atomic_and_x2"),
 557    (  -1,   -1,   -1, 0x69, 0x69, "s_buffer_atomic_or_x2"),
 558    (  -1,   -1,   -1, 0x6a, 0x6a, "s_buffer_atomic_xor_x2"),
 559    (  -1,   -1,   -1, 0x6b, 0x6b, "s_buffer_atomic_inc_x2"),
 560    (  -1,   -1,   -1, 0x6c, 0x6c, "s_buffer_atomic_dec_x2"),
 561    (  -1,   -1,   -1, 0x80, 0x80, "s_atomic_swap"),
 562    (  -1,   -1,   -1, 0x81, 0x81, "s_atomic_cmpswap"),
 563    (  -1,   -1,   -1, 0x82, 0x82, "s_atomic_add"),
 564    (  -1,   -1,   -1, 0x83, 0x83, "s_atomic_sub"),
 565    (  -1,   -1,   -1, 0x84, 0x84, "s_atomic_smin"),
 566    (  -1,   -1,   -1, 0x85, 0x85, "s_atomic_umin"),
 567    (  -1,   -1,   -1, 0x86, 0x86, "s_atomic_smax"),
 568    (  -1,   -1,   -1, 0x87, 0x87, "s_atomic_umax"),
 569    (  -1,   -1,   -1, 0x88, 0x88, "s_atomic_and"),
 570    (  -1,   -1,   -1, 0x89, 0x89, "s_atomic_or"),
 571    (  -1,   -1,   -1, 0x8a, 0x8a, "s_atomic_xor"),
 572    (  -1,   -1,   -1, 0x8b, 0x8b, "s_atomic_inc"),
 573    (  -1,   -1,   -1, 0x8c, 0x8c, "s_atomic_dec"),
 574    (  -1,   -1,   -1, 0xa0, 0xa0, "s_atomic_swap_x2"),
 575    (  -1,   -1,   -1, 0xa1, 0xa1, "s_atomic_cmpswap_x2"),
 576    (  -1,   -1,   -1, 0xa2, 0xa2, "s_atomic_add_x2"),
 577    (  -1,   -1,   -1, 0xa3, 0xa3, "s_atomic_sub_x2"),
 578    (  -1,   -1,   -1, 0xa4, 0xa4, "s_atomic_smin_x2"),
 579    (  -1,   -1,   -1, 0xa5, 0xa5, "s_atomic_umin_x2"),
 580    (  -1,   -1,   -1, 0xa6, 0xa6, "s_atomic_smax_x2"),
 581    (  -1,   -1,   -1, 0xa7, 0xa7, "s_atomic_umax_x2"),
 582    (  -1,   -1,   -1, 0xa8, 0xa8, "s_atomic_and_x2"),
 583    (  -1,   -1,   -1, 0xa9, 0xa9, "s_atomic_or_x2"),
 584    (  -1,   -1,   -1, 0xaa, 0xaa, "s_atomic_xor_x2"),
 585    (  -1,   -1,   -1, 0xab, 0xab, "s_atomic_inc_x2"),
 586    (  -1,   -1,   -1, 0xac, 0xac, "s_atomic_dec_x2"),
 587 }
 588 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in SMEM:
 589    opcode(name, gfx7, gfx9, gfx10, Format.SMEM, is_atomic = "atomic" in name)
 590
 591
 592 # VOP2 instructions: 2 inputs, 1 output (+ optional vcc)
 593 # TODO: misses some GFX6_7 opcodes which were shifted to VOP3 in GFX8
 594 VOP2 = {
 595   # GFX6, GFX7, GFX8, GFX9, GFX10, name, input/output modifiers
 596    (0x01, 0x01,   -1,   -1,   -1, "v_readlane_b32", False),
 597    (0x02, 0x02,   -1,   -1,   -1, "v_writelane_b32", False),
 598    (0x03, 0x03, 0x01, 0x01, 0x03, "v_add_f32", True),
 599    (0x04, 0x04, 0x02, 0x02, 0x04, "v_sub_f32", True),
 600    (0x05, 0x05, 0x03, 0x03, 0x05, "v_subrev_f32", True),
 601    (0x06, 0x06,   -1,   -1, 0x06, "v_mac_legacy_f32", True),
 602    (0x07, 0x07, 0x04, 0x04, 0x07, "v_mul_legacy_f32", True),
 603    (0x08, 0x08, 0x05, 0x05, 0x08, "v_mul_f32", True),
 604    (0x09, 0x09, 0x06, 0x06, 0x09, "v_mul_i32_i24", False),
 605    (0x0a, 0x0a, 0x07, 0x07, 0x0a, "v_mul_hi_i32_i24", False),
 606    (0x0b, 0x0b, 0x08, 0x08, 0x0b, "v_mul_u32_u24", False),
 607    (0x0c, 0x0c, 0x09, 0x09, 0x0c, "v_mul_hi_u32_u24", False),
 608    (0x0d, 0x0d,   -1,   -1,   -1, "v_min_legacy_f32", True),
 609    (0x0e, 0x0e,   -1,   -1,   -1, "v_max_legacy_f32", True),
 610    (0x0f, 0x0f, 0x0a, 0x0a, 0x0f, "v_min_f32", True),
 611    (0x10, 0x10, 0x0b, 0x0b, 0x10, "v_max_f32", True),
 612    (0x11, 0x11, 0x0c, 0x0c, 0x11, "v_min_i32", False),
 613    (0x12, 0x12, 0x0d, 0x0d, 0x12, "v_max_i32", False),
 614    (0x13, 0x13, 0x0e, 0x0e, 0x13, "v_min_u32", False),
 615    (0x14, 0x14, 0x0f, 0x0f, 0x14, "v_max_u32", False),
 616    (0x15, 0x15,   -1,   -1,   -1, "v_lshr_b32", False),
 617    (0x16, 0x16, 0x10, 0x10, 0x16, "v_lshrrev_b32", False),
 618    (0x17, 0x17,   -1,   -1,   -1, "v_ashr_i32", False),
 619    (0x18, 0x18, 0x11, 0x11, 0x18, "v_ashrrev_i32", False),
 620    (0x19, 0x19,   -1,   -1,   -1, "v_lshl_b32", False),
 621    (0x1a, 0x1a, 0x12, 0x12, 0x1a, "v_lshlrev_b32", False),
 622    (0x1b, 0x1b, 0x13, 0x13, 0x1b, "v_and_b32", False),
 623    (0x1c, 0x1c, 0x14, 0x14, 0x1c, "v_or_b32", False),
 624    (0x1d, 0x1d, 0x15, 0x15, 0x1d, "v_xor_b32", False),
 625    (  -1,   -1,   -1,   -1, 0x1e, "v_xnor_b32", False),
 626    (0x1f, 0x1f, 0x16, 0x16, 0x1f, "v_mac_f32", True),
 627    (0x20, 0x20, 0x17, 0x17, 0x20, "v_madmk_f32", False),
 628    (0x21, 0x21, 0x18, 0x18, 0x21, "v_madak_f32", False),
 629    (0x25, 0x25, 0x19, 0x19,   -1, "v_add_co_u32", False), # VOP3B only in RDNA
 630    (0x26, 0x26, 0x1a, 0x1a,   -1, "v_sub_co_u32", False), # VOP3B only in RDNA
 631    (0x27, 0x27, 0x1b, 0x1b,   -1, "v_subrev_co_u32", False), # VOP3B only in RDNA
 632    (0x28, 0x28, 0x1c, 0x1c, 0x28, "v_addc_co_u32", False), # v_add_co_ci_u32 in RDNA
 633    (0x29, 0x29, 0x1d, 0x1d, 0x29, "v_subb_co_u32", False), # v_sub_co_ci_u32 in RDNA
 634    (0x2a, 0x2a, 0x1e, 0x1e, 0x2a, "v_subbrev_co_u32", False), # v_subrev_co_ci_u32 in RDNA
 635    (  -1,   -1,   -1,   -1, 0x2b, "v_fmac_f32", True),
 636    (  -1,   -1,   -1,   -1, 0x2c, "v_fmamk_f32", True),
 637    (  -1,   -1,   -1,   -1, 0x2d, "v_fmaak_f32", True),
 638    (  -1,   -1, 0x1f, 0x1f, 0x32, "v_add_f16", True),
 639    (  -1,   -1, 0x20, 0x20, 0x33, "v_sub_f16", True),
 640    (  -1,   -1, 0x21, 0x21, 0x34, "v_subrev_f16", True),
 641    (  -1,   -1, 0x22, 0x22, 0x35, "v_mul_f16", True),
 642    (  -1,   -1, 0x23, 0x23,   -1, "v_mac_f16", True),
 643    (  -1,   -1, 0x24, 0x24,   -1, "v_madmk_f16", False),
 644    (  -1,   -1, 0x25, 0x25,   -1, "v_madak_f16", False),
 645    (  -1,   -1, 0x26, 0x26,   -1, "v_add_u16", False),
 646    (  -1,   -1, 0x27, 0x27,   -1, "v_sub_u16", False),
 647    (  -1,   -1, 0x28, 0x28,   -1, "v_subrev_u16", False),
 648    (  -1,   -1, 0x29, 0x29,   -1, "v_mul_lo_u16", False),
 649    (  -1,   -1, 0x2a, 0x2a,   -1, "v_lshlrev_b16", False),
 650    (  -1,   -1, 0x2b, 0x2b,   -1, "v_lshrrev_b16", False),
 651    (  -1,   -1, 0x2c, 0x2c,   -1, "v_ashrrev_b16", False),
 652    (  -1,   -1, 0x2d, 0x2d, 0x39, "v_max_f16", True),
 653    (  -1,   -1, 0x2e, 0x2e, 0x3a, "v_min_f16", True),
 654    (  -1,   -1, 0x2f, 0x2f,   -1, "v_max_u16", False),
 655    (  -1,   -1, 0x30, 0x30,   -1, "v_max_i16", False),
 656    (  -1,   -1, 0x31, 0x31,   -1, "v_min_u16", False),
 657    (  -1,   -1, 0x32, 0x32,   -1, "v_min_i16", False),
 658    (  -1,   -1, 0x33, 0x33, 0x3b, "v_ldexp_f16", False),
 659    (  -1,   -1, 0x34, 0x34, 0x25, "v_add_u32", False), # v_add_nc_u32 in RDNA
 660    (  -1,   -1, 0x35, 0x35, 0x26, "v_sub_u32", False), # v_sub_nc_u32 in RDNA
 661    (  -1,   -1, 0x36, 0x36, 0x27, "v_subrev_u32", False), # v_subrev_nc_u32 in RDNA
 662    (  -1,   -1,   -1,   -1, 0x36, "v_fmac_f16", False),
 663    (  -1,   -1,   -1,   -1, 0x37, "v_fmamk_f16", False),
 664    (  -1,   -1,   -1,   -1, 0x38, "v_fmaak_f16", False),
 665    (  -1,   -1,   -1,   -1, 0x3c, "v_pk_fmac_f16", False),
 666 }
 667 for (gfx6, gfx7, gfx8, gfx9, gfx10, name, modifiers) in VOP2:
 668    opcode(name, gfx7, gfx9, gfx10, Format.VOP2, modifiers, modifiers)
 669
 670 if True:
 671     # v_cndmask_b32 can use input modifiers but not output modifiers
 672     (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x00, 0x00, 0x00, 0x00, 0x01, "v_cndmask_b32")
 673     opcode(name, gfx7, gfx9, gfx10, Format.VOP2, True, False)
 674
 675
 676 # VOP1 instructions: instructions with 1 input and 1 output
 677 VOP1 = {
 678   # GFX6, GFX7, GFX8, GFX9, GFX10, name, input_modifiers, output_modifiers
 679    (0x00, 0x00, 0x00, 0x00, 0x00, "v_nop", False, False),
 680    (0x01, 0x01, 0x01, 0x01, 0x01, "v_mov_b32", False, False),
 681    (0x02, 0x02, 0x02, 0x02, 0x02, "v_readfirstlane_b32", False, False),
 682    (0x03, 0x03, 0x03, 0x03, 0x03, "v_cvt_i32_f64", True, False),
 683    (0x04, 0x04, 0x04, 0x04, 0x04, "v_cvt_f64_i32", False, True),
 684    (0x05, 0x05, 0x05, 0x05, 0x05, "v_cvt_f32_i32", False, True),
 685    (0x06, 0x06, 0x06, 0x06, 0x06, "v_cvt_f32_u32", False, True),
 686    (0x07, 0x07, 0x07, 0x07, 0x07, "v_cvt_u32_f32", True, False),
 687    (0x08, 0x08, 0x08, 0x08, 0x08, "v_cvt_i32_f32", True, False),
 688    (0x09, 0x09,   -1,   -1, 0x09, "v_mov_fed_b32", True, False), # LLVM mentions it for GFX8_9
 689    (0x0a, 0x0a, 0x0a, 0x0a, 0x0a, "v_cvt_f16_f32", True, True),
 690    (0x0b, 0x0b, 0x0b, 0x0b, 0x0b, "v_cvt_f32_f16", True, True),
 691    (0x0c, 0x0c, 0x0c, 0x0c, 0x0c, "v_cvt_rpi_i32_f32", True, False),
 692    (0x0d, 0x0d, 0x0d, 0x0d, 0x0d, "v_cvt_flr_i32_f32", True, False),
 693    (0x0e, 0x0e, 0x0e, 0x0e, 0x0e, "v_cvt_off_f32_i4", False, True),
 694    (0x0f, 0x0f, 0x0f, 0x0f, 0x0f, "v_cvt_f32_f64", True, True),
 695    (0x10, 0x10, 0x10, 0x10, 0x10, "v_cvt_f64_f32", True, True),
 696    (0x11, 0x11, 0x11, 0x11, 0x11, "v_cvt_f32_ubyte0", False, True),
 697    (0x12, 0x12, 0x12, 0x12, 0x12, "v_cvt_f32_ubyte1", False, True),
 698    (0x13, 0x13, 0x13, 0x13, 0x13, "v_cvt_f32_ubyte2", False, True),
 699    (0x14, 0x14, 0x14, 0x14, 0x14, "v_cvt_f32_ubyte3", False, True),
 700    (0x15, 0x15, 0x15, 0x15, 0x15, "v_cvt_u32_f64", True, False),
 701    (0x16, 0x16, 0x16, 0x16, 0x16, "v_cvt_f64_u32", False, True),
 702    (  -1, 0x17, 0x17, 0x17, 0x17, "v_trunc_f64", True, True),
 703    (  -1, 0x18, 0x18, 0x18, 0x18, "v_ceil_f64", True, True),
 704    (  -1, 0x19, 0x19, 0x19, 0x19, "v_rndne_f64", True, True),
 705    (  -1, 0x1a, 0x1a, 0x1a, 0x1a, "v_floor_f64", True, True),
 706    (  -1,   -1,   -1,   -1, 0x1b, "v_pipeflush", False, False),
 707    (0x20, 0x20, 0x1b, 0x1b, 0x20, "v_fract_f32", True, True),
 708    (0x21, 0x21, 0x1c, 0x1c, 0x21, "v_trunc_f32", True, True),
 709    (0x22, 0x22, 0x1d, 0x1d, 0x22, "v_ceil_f32", True, True),
 710    (0x23, 0x23, 0x1e, 0x1e, 0x23, "v_rndne_f32", True, True),
 711    (0x24, 0x24, 0x1f, 0x1f, 0x24, "v_floor_f32", True, True),
 712    (0x25, 0x25, 0x20, 0x20, 0x25, "v_exp_f32", True, True),
 713    (0x26, 0x26,   -1,   -1,   -1, "v_log_clamp_f32", True, True),
 714    (0x27, 0x27, 0x21, 0x21, 0x27, "v_log_f32", True, True),
 715    (0x28, 0x28,   -1,   -1,   -1, "v_rcp_clamp_f32", True, True),
 716    (0x29, 0x29,   -1,   -1,   -1, "v_rcp_legacy_f32", True, True),
 717    (0x2a, 0x2a, 0x22, 0x22, 0x2a, "v_rcp_f32", True, True),
 718    (0x2b, 0x2b, 0x23, 0x23, 0x2b, "v_rcp_iflag_f32", True, True),
 719    (0x2c, 0x2c,   -1,   -1,   -1, "v_rsq_clamp_f32", True, True),
 720    (0x2d, 0x2d,   -1,   -1,   -1, "v_rsq_legacy_f32", True, True),
 721    (0x2e, 0x2e, 0x24, 0x24, 0x2e, "v_rsq_f32", True, True),
 722    (0x2f, 0x2f, 0x25, 0x25, 0x2f, "v_rcp_f64", True, True),
 723    (0x30, 0x30,   -1,   -1,   -1, "v_rcp_clamp_f64", True, True),
 724    (0x31, 0x31, 0x26, 0x26, 0x31, "v_rsq_f64", True, True),
 725    (0x32, 0x32,   -1,   -1,   -1, "v_rsq_clamp_f64", True, True),
 726    (0x33, 0x33, 0x27, 0x27, 0x33, "v_sqrt_f32", True, True),
 727    (0x34, 0x34, 0x28, 0x28, 0x34, "v_sqrt_f64", True, True),
 728    (0x35, 0x35, 0x29, 0x29, 0x35, "v_sin_f32", True, True),
 729    (0x36, 0x36, 0x2a, 0x2a, 0x36, "v_cos_f32", True, True),
 730    (0x37, 0x37, 0x2b, 0x2b, 0x37, "v_not_b32", False, False),
 731    (0x38, 0x38, 0x2c, 0x2c, 0x38, "v_bfrev_b32", False, False),
 732    (0x39, 0x39, 0x2d, 0x2d, 0x39, "v_ffbh_u32", False, False),
 733    (0x3a, 0x3a, 0x2e, 0x2e, 0x3a, "v_ffbl_b32", False, False),
 734    (0x3b, 0x3b, 0x2f, 0x2f, 0x3b, "v_ffbh_i32", False, False),
 735    (0x3c, 0x3c, 0x30, 0x30, 0x3c, "v_frexp_exp_i32_f64", True, False),
 736    (0x3d, 0x3d, 0x31, 0x31, 0x3d, "v_frexp_mant_f64", True, False),
 737    (0x3e, 0x3e, 0x32, 0x32, 0x3e, "v_fract_f64", True, True),
 738    (0x3f, 0x3f, 0x33, 0x33, 0x3f, "v_frexp_exp_i32_f32", True, False),
 739    (0x40, 0x40, 0x34, 0x34, 0x40, "v_frexp_mant_f32", True, False),
 740    (0x41, 0x41, 0x35, 0x35, 0x41, "v_clrexcp", False, False),
 741    (0x42, 0x42, 0x36,   -1, 0x42, "v_movreld_b32", False, False),
 742    (0x43, 0x43, 0x37,   -1, 0x43, "v_movrels_b32", False, False),
 743    (0x44, 0x44, 0x38,   -1, 0x44, "v_movrelsd_b32", False, False),
 744    (  -1,   -1,   -1,   -1, 0x48, "v_movrelsd_2_b32", False, False),
 745    (  -1,   -1,   -1, 0x37,   -1, "v_screen_partition_4se_b32", False, False),
 746    (  -1,   -1, 0x39, 0x39, 0x50, "v_cvt_f16_u16", False, True),
 747    (  -1,   -1, 0x3a, 0x3a, 0x51, "v_cvt_f16_i16", False, True),
 748    (  -1,   -1, 0x3b, 0x3b, 0x52, "v_cvt_u16_f16", True, False),
 749    (  -1,   -1, 0x3c, 0x3c, 0x53, "v_cvt_i16_f16", True, False),
 750    (  -1,   -1, 0x3d, 0x3d, 0x54, "v_rcp_f16", True, True),
 751    (  -1,   -1, 0x3e, 0x3e, 0x55, "v_sqrt_f16", True, True),
 752    (  -1,   -1, 0x3f, 0x3f, 0x56, "v_rsq_f16", True, True),
 753    (  -1,   -1, 0x40, 0x40, 0x57, "v_log_f16", True, True),
 754    (  -1,   -1, 0x41, 0x41, 0x58, "v_exp_f16", True, True),
 755    (  -1,   -1, 0x42, 0x42, 0x59, "v_frexp_mant_f16", True, False),
 756    (  -1,   -1, 0x43, 0x43, 0x5a, "v_frexp_exp_i16_f16", True, False),
 757    (  -1,   -1, 0x44, 0x44, 0x5b, "v_floor_f16", True, True),
 758    (  -1,   -1, 0x45, 0x45, 0x5c, "v_ceil_f16", True, True),
 759    (  -1,   -1, 0x46, 0x46, 0x5d, "v_trunc_f16", True, True),
 760    (  -1,   -1, 0x47, 0x47, 0x5e, "v_rndne_f16", True, True),
 761    (  -1,   -1, 0x48, 0x48, 0x5f, "v_fract_f16", True, True),
 762    (  -1,   -1, 0x49, 0x49, 0x60, "v_sin_f16", True, True),
 763    (  -1,   -1, 0x4a, 0x4a, 0x61, "v_cos_f16", True, True),
 764    (  -1, 0x46, 0x4b, 0x4b,   -1, "v_exp_legacy_f32", True, True),
 765    (  -1, 0x45, 0x4c, 0x4c,   -1, "v_log_legacy_f32", True, True),
 766    (  -1,   -1,   -1, 0x4f, 0x62, "v_sat_pk_u8_i16", False, False),
 767    (  -1,   -1,   -1, 0x4d, 0x63, "v_cvt_norm_i16_f16", True, False),
 768    (  -1,   -1,   -1, 0x4e, 0x64, "v_cvt_norm_u16_f16", True, False),
 769    (  -1,   -1,   -1, 0x51, 0x65, "v_swap_b32", False, False),
 770    (  -1,   -1,   -1,   -1, 0x68, "v_swaprel_b32", False, False),
 771 }
 772 for (gfx6, gfx7, gfx8, gfx9, gfx10, name, in_mod, out_mod) in VOP1:
 773    opcode(name, gfx7, gfx9, gfx10, Format.VOP1, in_mod, out_mod)
 774
 775
 776 # VOPC instructions:
 777
 778 VOPC_CLASS = {
 779    (0x88, 0x88, 0x10, 0x10, 0x88, "v_cmp_class_f32"),
 780    (  -1,   -1, 0x14, 0x14, 0x8f, "v_cmp_class_f16"),
 781    (0x98, 0x98, 0x11, 0x11, 0x98, "v_cmpx_class_f32"),
 782    (  -1,   -1, 0x15, 0x15, 0x9f, "v_cmpx_class_f16"),
 783    (0xa8, 0xa8, 0x12, 0x12, 0xa8, "v_cmp_class_f64"),
 784    (0xb8, 0xb8, 0x13, 0x13, 0xb8, "v_cmpx_class_f64"),
 785 }
 786 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in VOPC_CLASS:
 787     opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 788
 789 COMPF = ["f", "lt", "eq", "le", "gt", "lg", "ge", "o", "u", "nge", "nlg", "ngt", "nle", "neq", "nlt", "tru"]
 790
 791 for i in range(8):
 792    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0x20+i, 0x20+i, 0xc8+i, "v_cmp_"+COMPF[i]+"_f16")
 793    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 794    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0x30+i, 0x30+i, 0xd8+i, "v_cmpx_"+COMPF[i]+"_f16")
 795    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 796    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0x28+i, 0x28+i, 0xe8+i, "v_cmp_"+COMPF[i+8]+"_f16")
 797    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 798    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0x38+i, 0x38+i, 0xf8+i, "v_cmpx_"+COMPF[i+8]+"_f16")
 799    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 800
 801 for i in range(16):
 802    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x00+i, 0x00+i, 0x40+i, 0x40+i, 0x00+i, "v_cmp_"+COMPF[i]+"_f32")
 803    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 804    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x10+i, 0x10+i, 0x50+i, 0x50+i, 0x10+i, "v_cmpx_"+COMPF[i]+"_f32")
 805    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 806    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x20+i, 0x20+i, 0x60+i, 0x60+i, 0x20+i, "v_cmp_"+COMPF[i]+"_f64")
 807    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 808    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x30+i, 0x30+i, 0x70+i, 0x70+i, 0x30+i, "v_cmpx_"+COMPF[i]+"_f64")
 809    opcode(name, gfx7, gfx9, gfx10, Format.VOPC, True, False)
 810    # GFX_6_7
 811    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x40+i, 0x40+i, -1, -1, -1, "v_cmps_"+COMPF[i]+"_f32")
 812    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x50+i, 0x50+i, -1, -1, -1, "v_cmpsx_"+COMPF[i]+"_f32")
 813    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x60+i, 0x60+i, -1, -1, -1, "v_cmps_"+COMPF[i]+"_f64")
 814    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x70+i, 0x70+i, -1, -1, -1, "v_cmpsx_"+COMPF[i]+"_f64")
 815
 816 COMPI = ["f", "lt", "eq", "le", "gt", "lg", "ge", "tru"]
 817
 818 # GFX_8_9
 819 for i in [0,7]: # only 0 and 7
 820    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xa0+i, 0xa0+i, -1, "v_cmp_"+COMPI[i]+"_i16")
 821    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 822    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xb0+i, 0xb0+i, -1, "v_cmpx_"+COMPI[i]+"_i16")
 823    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 824    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xa8+i, 0xa8+i, -1, "v_cmp_"+COMPI[i]+"_u16")
 825    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 826    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xb8+i, 0xb8+i, -1, "v_cmpx_"+COMPI[i]+"_u16")
 827    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 828
 829 for i in range(1, 7): # [1..6]
 830    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xa0+i, 0xa0+i, 0x88+i, "v_cmp_"+COMPI[i]+"_i16")
 831    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 832    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xb0+i, 0xb0+i, 0x98+i, "v_cmpx_"+COMPI[i]+"_i16")
 833    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 834    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xa8+i, 0xa8+i, 0xa8+i, "v_cmp_"+COMPI[i]+"_u16")
 835    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 836    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, 0xb8+i, 0xb8+i, 0xb8+i, "v_cmpx_"+COMPI[i]+"_u16")
 837    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 838
 839 for i in range(8):
 840    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x80+i, 0x80+i, 0xc0+i, 0xc0+i, 0x80+i, "v_cmp_"+COMPI[i]+"_i32")
 841    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 842    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0x90+i, 0x90+i, 0xd0+i, 0xd0+i, 0x90+i, "v_cmpx_"+COMPI[i]+"_i32")
 843    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 844    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0xa0+i, 0xa0+i, 0xe0+i, 0xe0+i, 0xa0+i, "v_cmp_"+COMPI[i]+"_i64")
 845    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 846    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0xb0+i, 0xb0+i, 0xf0+i, 0xf0+i, 0xb0+i, "v_cmpx_"+COMPI[i]+"_i64")
 847    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 848    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0xc0+i, 0xc0+i, 0xc8+i, 0xc8+i, 0xc0+i, "v_cmp_"+COMPI[i]+"_u32")
 849    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 850    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0xd0+i, 0xd0+i, 0xd8+i, 0xd8+i, 0xd0+i, "v_cmpx_"+COMPI[i]+"_u32")
 851    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 852    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0xe0+i, 0xe0+i, 0xe8+i, 0xe8+i, 0xe0+i, "v_cmp_"+COMPI[i]+"_u64")
 853    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 854    (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (0xf0+i, 0xf0+i, 0xf8+i, 0xf8+i, 0xf0+i, "v_cmpx_"+COMPI[i]+"_u64")
 855    opcode(name, gfx7, gfx9, gfx10, Format.VOPC)
 856
 857
 858 # VOPP instructions: packed 16bit instructions - 1 or 2 inputs and 1 output
 859 VOPP = {
 860    (0x00, "v_pk_mad_i16"),
 861    (0x01, "v_pk_mul_lo_u16"),
 862    (0x02, "v_pk_add_i16"),
 863    (0x03, "v_pk_sub_i16"),
 864    (0x04, "v_pk_lshlrev_b16"),
 865    (0x05, "v_pk_lshrrev_b16"),
 866    (0x06, "v_pk_ashrrev_i16"),
 867    (0x07, "v_pk_max_i16"),
 868    (0x08, "v_pk_min_i16"),
 869    (0x09, "v_pk_mad_u16"),
 870    (0x0a, "v_pk_add_u16"),
 871    (0x0b, "v_pk_sub_u16"),
 872    (0x0c, "v_pk_max_u16"),
 873    (0x0d, "v_pk_min_u16"),
 874    (0x0e, "v_pk_fma_f16"),
 875    (0x0f, "v_pk_add_f16"),
 876    (0x10, "v_pk_mul_f16"),
 877    (0x11, "v_pk_min_f16"),
 878    (0x12, "v_pk_max_f16"),
 879    (0x20, "v_pk_fma_mix_f32"), # v_mad_mix_f32 in VEGA ISA, v_fma_mix_f32 in RDNA ISA
 880    (0x21, "v_pk_fma_mixlo_f16"), # v_mad_mixlo_f16 in VEGA ISA, v_fma_mixlo_f16 in RDNA ISA
 881    (0x22, "v_pk_fma_mixhi_f16"), # v_mad_mixhi_f16 in VEGA ISA, v_fma_mixhi_f16 in RDNA ISA
 882 }
 883 # note that these are only supported on gfx9+ so we'll need to distinguish between gfx8 and gfx9 here
 884 # (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (-1, -1, -1, code, code, name)
 885 for (code, name) in VOPP:
 886    opcode(name, -1, code, code, Format.VOP3P)
 887
 888
 889 # VINTERP instructions:
 890 VINTRP = {
 891    (0x00, "v_interp_p1_f32"),
 892    (0x01, "v_interp_p2_f32"),
 893    (0x02, "v_interp_mov_f32"),
 894 }
 895 # (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (code, code, code, code, code, name)
 896 for (code, name) in VINTRP:
 897    opcode(name, code, code, code, Format.VINTRP)
 898
 899 # VOP3 instructions: 3 inputs, 1 output
 900 # VOP3b instructions: have a unique scalar output, e.g. VOP2 with vcc out
 901 VOP3 = {
 902    (0x140, 0x140, 0x1c0, 0x1c0, 0x140, "v_mad_legacy_f32", True, True),
 903    (0x141, 0x141, 0x1c1, 0x1c1, 0x141, "v_mad_f32", True, True),
 904    (0x142, 0x142, 0x1c2, 0x1c2, 0x142, "v_mad_i32_i24", False, False),
 905    (0x143, 0x143, 0x1c3, 0x1c3, 0x143, "v_mad_u32_u24", False, False),
 906    (0x144, 0x144, 0x1c4, 0x1c4, 0x144, "v_cubeid_f32", True, True),
 907    (0x145, 0x145, 0x1c5, 0x1c5, 0x145, "v_cubesc_f32", True, True),
 908    (0x146, 0x146, 0x1c6, 0x1c6, 0x146, "v_cubetc_f32", True, True),
 909    (0x147, 0x147, 0x1c7, 0x1c7, 0x147, "v_cubema_f32", True, True),
 910    (0x148, 0x148, 0x1c8, 0x1c8, 0x148, "v_bfe_u32", False, False),
 911    (0x149, 0x149, 0x1c9, 0x1c9, 0x149, "v_bfe_i32", False, False),
 912    (0x14a, 0x14a, 0x1ca, 0x1ca, 0x14a, "v_bfi_b32", False, False),
 913    (0x14b, 0x14b, 0x1cb, 0x1cb, 0x14b, "v_fma_f32", True, True),
 914    (0x14c, 0x14c, 0x1cc, 0x1cc, 0x14c, "v_fma_f64", True, True),
 915    (0x14d, 0x14d, 0x1cd, 0x1cd, 0x14d, "v_lerp_u8", False, False),
 916    (0x14e, 0x14e, 0x1ce, 0x1ce, 0x14e, "v_alignbit_b32", False, False),
 917    (0x14f, 0x14f, 0x1cf, 0x1cf, 0x14f, "v_alignbyte_b32", False, False),
 918    (0x150, 0x150,    -1,    -1, 0x150, "v_mullit_f32", True, True),
 919    (0x151, 0x151, 0x1d0, 0x1d0, 0x151, "v_min3_f32", True, True),
 920    (0x152, 0x152, 0x1d1, 0x1d1, 0x152, "v_min3_i32", False, False),
 921    (0x153, 0x153, 0x1d2, 0x1d2, 0x153, "v_min3_u32", False, False),
 922    (0x154, 0x154, 0x1d3, 0x1d3, 0x154, "v_max3_f32", True, True),
 923    (0x155, 0x155, 0x1d4, 0x1d4, 0x155, "v_max3_i32", False, False),
 924    (0x156, 0x156, 0x1d5, 0x1d5, 0x156, "v_max3_u32", False, False),
 925    (0x157, 0x157, 0x1d6, 0x1d6, 0x157, "v_med3_f32", True, True),
 926    (0x158, 0x158, 0x1d7, 0x1d7, 0x158, "v_med3_i32", False, False),
 927    (0x159, 0x159, 0x1d8, 0x1d8, 0x159, "v_med3_u32", False, False),
 928    (0x15a, 0x15a, 0x1d9, 0x1d9, 0x15a, "v_sad_u8", False, False),
 929    (0x15b, 0x15b, 0x1da, 0x1da, 0x15b, "v_sad_hi_u8", False, False),
 930    (0x15c, 0x15c, 0x1db, 0x1db, 0x15c, "v_sad_u16", False, False),
 931    (0x15d, 0x15d, 0x1dc, 0x1dc, 0x15d, "v_sad_u32", False, False),
 932    (0x15e, 0x15e, 0x1dd, 0x1dd, 0x15e, "v_cvt_pk_u8_f32", True, False),
 933    (0x15f, 0x15f, 0x1de, 0x1de, 0x15f, "v_div_fixup_f32", True, True),
 934    (0x160, 0x160, 0x1df, 0x1df, 0x160, "v_div_fixup_f64", True, True),
 935    (0x161, 0x161,    -1,    -1,    -1, "v_lshl_b64", False, False),
 936    (0x162, 0x162,    -1,    -1,    -1, "v_lshr_b64", False, False),
 937    (0x163, 0x163,    -1,    -1,    -1, "v_ashr_i64", False, False),
 938    (0x164, 0x164, 0x280, 0x280, 0x164, "v_add_f64", True, True),
 939    (0x165, 0x165, 0x281, 0x281, 0x165, "v_mul_f64", True, True),
 940    (0x166, 0x166, 0x282, 0x282, 0x166, "v_min_f64", True, True),
 941    (0x167, 0x167, 0x283, 0x283, 0x167, "v_max_f64", True, True),
 942    (0x168, 0x168, 0x284, 0x284, 0x168, "v_ldexp_f64", False, True), # src1 can take input modifiers
 943    (0x169, 0x169, 0x285, 0x285, 0x169, "v_mul_lo_u32", False, False),
 944    (0x16a, 0x16a, 0x286, 0x286, 0x16a, "v_mul_hi_u32", False, False),
 945    (0x16b, 0x16b, 0x285, 0x285, 0x16b, "v_mul_lo_i32", False, False), # identical to v_mul_lo_u32
 946    (0x16c, 0x16c, 0x287, 0x287, 0x16c, "v_mul_hi_i32", False, False),
 947    (0x16d, 0x16d, 0x1e0, 0x1e0, 0x16d, "v_div_scale_f32", True, True), # writes to VCC
 948    (0x16e, 0x16e, 0x1e1, 0x1e1, 0x16e, "v_div_scale_f64", True, True), # writes to VCC
 949    (0x16f, 0x16f, 0x1e2, 0x1e2, 0x16f, "v_div_fmas_f32", True, True), # takes VCC input
 950    (0x170, 0x170, 0x1e3, 0x1e3, 0x170, "v_div_fmas_f64", True, True), # takes VCC input
 951    (0x171, 0x171, 0x1e4, 0x1e4, 0x171, "v_msad_u8", False, False),
 952    (0x172, 0x172, 0x1e5, 0x1e5, 0x172, "v_qsad_pk_u16_u8", False, False),
 953    (0x172,    -1,    -1,    -1,    -1, "v_qsad_u8", False, False), # what's the difference?
 954    (0x173, 0x173, 0x1e6, 0x1e6, 0x173, "v_mqsad_pk_u16_u8", False, False),
 955    (0x173,    -1,    -1,    -1,    -1, "v_mqsad_u8", False, False), # what's the difference?
 956    (0x174, 0x174, 0x292, 0x292, 0x174, "v_trig_preop_f64", False, False),
 957    (   -1, 0x175, 0x1e7, 0x1e7, 0x175, "v_mqsad_u32_u8", False, False),
 958    (   -1, 0x176, 0x1e8, 0x1e8, 0x176, "v_mad_u64_u32", False, False),
 959    (   -1, 0x177, 0x1e9, 0x1e9, 0x177, "v_mad_i64_i32", False, False),
 960    (   -1,    -1, 0x1ea, 0x1ea,    -1, "v_mad_legacy_f16", True, True),
 961    (   -1,    -1, 0x1eb, 0x1eb,    -1, "v_mad_legacy_u16", False, False),
 962    (   -1,    -1, 0x1ec, 0x1ec,    -1, "v_mad_legacy_i16", False, False),
 963    (   -1,    -1, 0x1ed, 0x1ed, 0x344, "v_perm_b32", False, False),
 964    (   -1,    -1, 0x1ee, 0x1ee,    -1, "v_fma_legacy_f16", True, True),
 965    (   -1,    -1, 0x1ef, 0x1ef,    -1, "v_div_fixup_legacy_f16", True, True),
 966    (0x12c, 0x12c, 0x1f0, 0x1f0,    -1, "v_cvt_pkaccum_u8_f32", True, False),
 967    (   -1,    -1,    -1, 0x1f1, 0x373, "v_mad_u32_u16", False, False),
 968    (   -1,    -1,    -1, 0x1f2, 0x375, "v_mad_i32_i16", False, False),
 969    (   -1,    -1,    -1, 0x1f3, 0x345, "v_xad_u32", False, False),
 970    (   -1,    -1,    -1, 0x1f4, 0x351, "v_min3_f16", True, True),
 971    (   -1,    -1,    -1, 0x1f5, 0x352, "v_min3_i16", False, False),
 972    (   -1,    -1,    -1, 0x1f6, 0x353, "v_min3_u16", False, False),
 973    (   -1,    -1,    -1, 0x1f7, 0x354, "v_max3_f16", True, True),
 974    (   -1,    -1,    -1, 0x1f8, 0x355, "v_max3_i16", False, False),
 975    (   -1,    -1,    -1, 0x1f9, 0x356, "v_max3_u16", False, False),
 976    (   -1,    -1,    -1, 0x1fa, 0x357, "v_med3_f16", True, True),
 977    (   -1,    -1,    -1, 0x1fb, 0x358, "v_med3_i16", False, False),
 978    (   -1,    -1,    -1, 0x1fc, 0x359, "v_med3_u16", False, False),
 979    (   -1,    -1,    -1, 0x1fd, 0x346, "v_lshl_add_u32", False, False),
 980    (   -1,    -1,    -1, 0x1fe, 0x347, "v_add_lshl_u32", False, False),
 981    (   -1,    -1,    -1, 0x1ff, 0x36d, "v_add3_u32", False, False),
 982    (   -1,    -1,    -1, 0x200, 0x36f, "v_lshl_or_b32", False, False),
 983    (   -1,    -1,    -1, 0x201, 0x371, "v_and_or_b32", False, False),
 984    (   -1,    -1,    -1, 0x202, 0x372, "v_or3_b32", False, False),
 985    (   -1,    -1,    -1, 0x203,    -1, "v_mad_f16", True, True),
 986    (   -1,    -1,    -1, 0x204, 0x340, "v_mad_u16", False, False),
 987    (   -1,    -1,    -1, 0x205, 0x35e, "v_mad_i16", False, False),
 988    (   -1,    -1,    -1, 0x206, 0x34b, "v_fma_f16", True, True),
 989    (   -1,    -1,    -1, 0x207, 0x35f, "v_div_fixup_f16", True, True),
 990    (   -1,    -1, 0x274, 0x274, 0x342, "v_interp_p1ll_f16", True, True),
 991    (   -1,    -1, 0x275, 0x275, 0x343, "v_interp_p1lv_f16", True, True),
 992    (   -1,    -1, 0x276, 0x276,    -1, "v_interp_p2_legacy_f16", True, True),
 993    (   -1,    -1,    -1, 0x277, 0x35a, "v_interp_p2_f16", True, True),
 994    (0x12b, 0x12b, 0x288, 0x288, 0x362, "v_ldexp_f32", False, True),
 995    (   -1,    -1, 0x289, 0x289, 0x360, "v_readlane_b32_e64", False, False),
 996    (   -1,    -1, 0x28a, 0x28a, 0x361, "v_writelane_b32_e64", False, False),
 997    (0x122, 0x122, 0x28b, 0x28b, 0x364, "v_bcnt_u32_b32", False, False),
 998    (0x123, 0x123, 0x28c, 0x28c, 0x365, "v_mbcnt_lo_u32_b32", False, False),
 999    (0x124, 0x124, 0x28d, 0x28d, 0x366, "v_mbcnt_hi_u32_b32", False, False),
1000    (   -1,    -1, 0x28f, 0x28f, 0x2ff, "v_lshlrev_b64", False, False),
1001    (   -1,    -1, 0x290, 0x290, 0x300, "v_lshrrev_b64", False, False),
1002    (   -1,    -1, 0x291, 0x291, 0x301, "v_ashrrev_i64", False, False),
1003    (0x11e, 0x11e, 0x293, 0x293, 0x363, "v_bfm_b32", False, False),
1004    (0x12d, 0x12d, 0x294, 0x294, 0x368, "v_cvt_pknorm_i16_f32", True, False),
1005    (0x12e, 0x12e, 0x295, 0x295, 0x369, "v_cvt_pknorm_u16_f32", True, False),
1006    (0x12f, 0x12f, 0x296, 0x296, 0x12f, "v_cvt_pkrtz_f16_f32", True, False), # GFX6_7_10 is VOP2 with opcode 0x02f
1007    (0x130, 0x130, 0x297, 0x297, 0x36a, "v_cvt_pk_u16_u32", False, False),
1008    (0x131, 0x131, 0x298, 0x298, 0x36b, "v_cvt_pk_i16_i32", False, False),
1009    (   -1,    -1,    -1, 0x299, 0x312, "v_cvt_pknorm_i16_f16", True, False),
1010    (   -1,    -1,    -1, 0x29a, 0x313, "v_cvt_pknorm_u16_f16", True, False),
1011    (   -1,    -1,    -1, 0x29c, 0x37f, "v_add_i32", False, False),
1012    (   -1,    -1,    -1, 0x29d, 0x376, "v_sub_i32", False, False),
1013    (   -1,    -1,    -1, 0x29e, 0x30d, "v_add_i16", False, False),
1014    (   -1,    -1,    -1, 0x29f, 0x30e, "v_sub_i16", False, False),
1015    (   -1,    -1,    -1, 0x2a0, 0x311, "v_pack_b32_f16", True, False),
1016    (   -1,    -1,    -1,    -1, 0x178, "v_xor3_b32", False, False),
1017    (   -1,    -1,    -1,    -1, 0x377, "v_permlane16_b32", False, False),
1018    (   -1,    -1,    -1,    -1, 0x378, "v_permlanex16_b32", False, False),
1019    (   -1,    -1,    -1,    -1, 0x30f, "v_add_co_u32_e64", False, False),
1020    (   -1,    -1,    -1,    -1, 0x310, "v_sub_co_u32_e64", False, False),
1021    (   -1,    -1,    -1,    -1, 0x319, "v_subrev_co_u32_e64", False, False),
1022 # TODO: many 16bit instructions moved from VOP2 to VOP3 on GFX10
1023 }
1024 for (gfx6, gfx7, gfx8, gfx9, gfx10, name, in_mod, out_mod) in VOP3:
1025    opcode(name, gfx7, gfx9, gfx10, Format.VOP3A, in_mod, out_mod)
1026
1027
1028 # DS instructions: 3 inputs (1 addr, 2 data), 1 output
1029 DS = {
1030    (0x00, 0x00, 0x00, 0x00, 0x00, "ds_add_u32"),
1031    (0x01, 0x01, 0x01, 0x01, 0x01, "ds_sub_u32"),
1032    (0x02, 0x02, 0x02, 0x02, 0x02, "ds_rsub_u32"),
1033    (0x03, 0x03, 0x03, 0x03, 0x03, "ds_inc_u32"),
1034    (0x04, 0x04, 0x04, 0x04, 0x04, "ds_dec_u32"),
1035    (0x05, 0x05, 0x05, 0x05, 0x05, "ds_min_i32"),
1036    (0x06, 0x06, 0x06, 0x06, 0x06, "ds_max_i32"),
1037    (0x07, 0x07, 0x07, 0x07, 0x07, "ds_min_u32"),
1038    (0x08, 0x08, 0x08, 0x08, 0x08, "ds_max_u32"),
1039    (0x09, 0x09, 0x09, 0x09, 0x09, "ds_and_b32"),
1040    (0x0a, 0x0a, 0x0a, 0x0a, 0x0a, "ds_or_b32"),
1041    (0x0b, 0x0b, 0x0b, 0x0b, 0x0b, "ds_xor_b32"),
1042    (0x0c, 0x0c, 0x0c, 0x0c, 0x0c, "ds_mskor_b32"),
1043    (0x0d, 0x0d, 0x0d, 0x0d, 0x0d, "ds_write_b32"),
1044    (0x0e, 0x0e, 0x0e, 0x0e, 0x0e, "ds_write2_b32"),
1045    (0x0f, 0x0f, 0x0f, 0x0f, 0x0f, "ds_write2st64_b32"),
1046    (0x10, 0x10, 0x10, 0x10, 0x10, "ds_cmpst_b32"),
1047    (0x11, 0x11, 0x11, 0x11, 0x11, "ds_cmpst_f32"),
1048    (0x12, 0x12, 0x12, 0x12, 0x12, "ds_min_f32"),
1049    (0x13, 0x13, 0x13, 0x13, 0x13, "ds_max_f32"),
1050    (  -1, 0x14, 0x14, 0x14, 0x14, "ds_nop"),
1051    (  -1,   -1, 0x15, 0x15, 0x15, "ds_add_f32"),
1052    (  -1,   -1, 0x1d, 0x1d, 0xb0, "ds_write_addtid_b32"),
1053    (0x1e, 0x1e, 0x1e, 0x1e, 0x1e, "ds_write_b8"),
1054    (0x1f, 0x1f, 0x1f, 0x1f, 0x1f, "ds_write_b16"),
1055    (0x20, 0x20, 0x20, 0x20, 0x20, "ds_add_rtn_u32"),
1056    (0x21, 0x21, 0x21, 0x21, 0x21, "ds_sub_rtn_u32"),
1057    (0x22, 0x22, 0x22, 0x22, 0x22, "ds_rsub_rtn_u32"),
1058    (0x23, 0x23, 0x23, 0x23, 0x23, "ds_inc_rtn_u32"),
1059    (0x24, 0x24, 0x24, 0x24, 0x24, "ds_dec_rtn_u32"),
1060    (0x25, 0x25, 0x25, 0x25, 0x25, "ds_min_rtn_i32"),
1061    (0x26, 0x26, 0x26, 0x26, 0x26, "ds_max_rtn_i32"),
1062    (0x27, 0x27, 0x27, 0x27, 0x27, "ds_min_rtn_u32"),
1063    (0x28, 0x28, 0x28, 0x28, 0x28, "ds_max_rtn_u32"),
1064    (0x29, 0x29, 0x29, 0x29, 0x29, "ds_and_rtn_b32"),
1065    (0x2a, 0x2a, 0x2a, 0x2a, 0x2a, "ds_or_rtn_b32"),
1066    (0x2b, 0x2b, 0x2b, 0x2b, 0x2b, "ds_xor_rtn_b32"),
1067    (0x2c, 0x2c, 0x2c, 0x2c, 0x2c, "ds_mskor_rtn_b32"),
1068    (0x2d, 0x2d, 0x2d, 0x2d, 0x2d, "ds_wrxchg_rtn_b32"),
1069    (0x2e, 0x2e, 0x2e, 0x2e, 0x2e, "ds_wrxchg2_rtn_b32"),
1070    (0x2f, 0x2f, 0x2f, 0x2f, 0x2f, "ds_wrxchg2st64_rtn_b32"),
1071    (0x30, 0x30, 0x30, 0x30, 0x30, "ds_cmpst_rtn_b32"),
1072    (0x31, 0x31, 0x31, 0x31, 0x31, "ds_cmpst_rtn_f32"),
1073    (0x32, 0x32, 0x32, 0x32, 0x32, "ds_min_rtn_f32"),
1074    (0x33, 0x33, 0x33, 0x33, 0x33, "ds_max_rtn_f32"),
1075    (  -1, 0x34, 0x34, 0x34, 0x34, "ds_wrap_rtn_b32"),
1076    (  -1,   -1, 0x35, 0x35, 0x55, "ds_add_rtn_f32"),
1077    (0x36, 0x36, 0x36, 0x36, 0x36, "ds_read_b32"),
1078    (0x37, 0x37, 0x37, 0x37, 0x37, "ds_read2_b32"),
1079    (0x38, 0x38, 0x38, 0x38, 0x38, "ds_read2st64_b32"),
1080    (0x39, 0x39, 0x39, 0x39, 0x39, "ds_read_i8"),
1081    (0x3a, 0x3a, 0x3a, 0x3a, 0x3a, "ds_read_u8"),
1082    (0x3b, 0x3b, 0x3b, 0x3b, 0x3b, "ds_read_i16"),
1083    (0x3c, 0x3c, 0x3c, 0x3c, 0x3c, "ds_read_u16"),
1084    (0x35, 0x35, 0x3d, 0x3d, 0x35, "ds_swizzle_b32"), #data1 & offset, no addr/data2
1085    (  -1,   -1, 0x3e, 0x3e, 0xb2, "ds_permute_b32"),
1086    (  -1,   -1, 0x3f, 0x3f, 0xb3, "ds_bpermute_b32"),
1087    (0x40, 0x40, 0x40, 0x40, 0x40, "ds_add_u64"),
1088    (0x41, 0x41, 0x41, 0x41, 0x41, "ds_sub_u64"),
1089    (0x42, 0x42, 0x42, 0x42, 0x42, "ds_rsub_u64"),
1090    (0x43, 0x43, 0x43, 0x43, 0x43, "ds_inc_u64"),
1091    (0x44, 0x44, 0x44, 0x44, 0x44, "ds_dec_u64"),
1092    (0x45, 0x45, 0x45, 0x45, 0x45, "ds_min_i64"),
1093    (0x46, 0x46, 0x46, 0x46, 0x46, "ds_max_i64"),
1094    (0x47, 0x47, 0x47, 0x47, 0x47, "ds_min_u64"),
1095    (0x48, 0x48, 0x48, 0x48, 0x48, "ds_max_u64"),
1096    (0x49, 0x49, 0x49, 0x49, 0x49, "ds_and_b64"),
1097    (0x4a, 0x4a, 0x4a, 0x4a, 0x4a, "ds_or_b64"),
1098    (0x4b, 0x4b, 0x4b, 0x4b, 0x4b, "ds_xor_b64"),
1099    (0x4c, 0x4c, 0x4c, 0x4c, 0x4c, "ds_mskor_b64"),
1100    (0x4d, 0x4d, 0x4d, 0x4d, 0x4d, "ds_write_b64"),
1101    (0x4e, 0x4e, 0x4e, 0x4e, 0x4e, "ds_write2_b64"),
1102    (0x4f, 0x4f, 0x4f, 0x4f, 0x4f, "ds_write2st64_b64"),
1103    (0x50, 0x50, 0x50, 0x50, 0x50, "ds_cmpst_b64"),
1104    (0x51, 0x51, 0x51, 0x51, 0x51, "ds_cmpst_f64"),
1105    (0x52, 0x52, 0x52, 0x52, 0x52, "ds_min_f64"),
1106    (0x53, 0x53, 0x53, 0x53, 0x53, "ds_max_f64"),
1107    (  -1,   -1, 0x54, 0x54, 0xa0, "ds_write_b8_d16_hi"),
1108    (  -1,   -1, 0x55, 0x55, 0xa1, "ds_write_b16_d16_hi"),
1109    (  -1,   -1, 0x56, 0x56, 0xa2, "ds_read_u8_d16"),
1110    (  -1,   -1, 0x57, 0x57, 0xa3, "ds_read_u8_d16_hi"),
1111    (  -1,   -1, 0x58, 0x58, 0xa4, "ds_read_i8_d16"),
1112    (  -1,   -1, 0x59, 0x59, 0xa5, "ds_read_i8_d16_hi"),
1113    (  -1,   -1, 0x5a, 0x5a, 0xa6, "ds_read_u16_d16"),
1114    (  -1,   -1, 0x5b, 0x5b, 0xa7, "ds_read_u16_d16_hi"),
1115    (0x60, 0x60, 0x60, 0x60, 0x60, "ds_add_rtn_u64"),
1116    (0x61, 0x61, 0x61, 0x61, 0x61, "ds_sub_rtn_u64"),
1117    (0x62, 0x62, 0x62, 0x62, 0x62, "ds_rsub_rtn_u64"),
1118    (0x63, 0x63, 0x63, 0x63, 0x63, "ds_inc_rtn_u64"),
1119    (0x64, 0x64, 0x64, 0x64, 0x64, "ds_dec_rtn_u64"),
1120    (0x65, 0x65, 0x65, 0x65, 0x65, "ds_min_rtn_i64"),
1121    (0x66, 0x66, 0x66, 0x66, 0x66, "ds_max_rtn_i64"),
1122    (0x67, 0x67, 0x67, 0x67, 0x67, "ds_min_rtn_u64"),
1123    (0x68, 0x68, 0x68, 0x68, 0x68, "ds_max_rtn_u64"),
1124    (0x69, 0x69, 0x69, 0x69, 0x69, "ds_and_rtn_b64"),
1125    (0x6a, 0x6a, 0x6a, 0x6a, 0x6a, "ds_or_rtn_b64"),
1126    (0x6b, 0x6b, 0x6b, 0x6b, 0x6b, "ds_xor_rtn_b64"),
1127    (0x6c, 0x6c, 0x6c, 0x6c, 0x6c, "ds_mskor_rtn_b64"),
1128    (0x6d, 0x6d, 0x6d, 0x6d, 0x6d, "ds_wrxchg_rtn_b64"),
1129    (0x6e, 0x6e, 0x6e, 0x6e, 0x6e, "ds_wrxchg2_rtn_b64"),
1130    (0x6f, 0x6f, 0x6f, 0x6f, 0x6f, "ds_wrxchg2st64_rtn_b64"),
1131    (0x70, 0x70, 0x70, 0x70, 0x70, "ds_cmpst_rtn_b64"),
1132    (0x71, 0x71, 0x71, 0x71, 0x71, "ds_cmpst_rtn_f64"),
1133    (0x72, 0x72, 0x72, 0x72, 0x72, "ds_min_rtn_f64"),
1134    (0x73, 0x73, 0x73, 0x73, 0x73, "ds_max_rtn_f64"),
1135    (0x76, 0x76, 0x76, 0x76, 0x76, "ds_read_b64"),
1136    (0x77, 0x77, 0x77, 0x77, 0x77, "ds_read2_b64"),
1137    (0x78, 0x78, 0x78, 0x78, 0x78, "ds_read2st64_b64"),
1138    (  -1, 0x7e, 0x7e, 0x7e, 0x7e, "ds_condxchg32_rtn_b64"),
1139    (0x80, 0x80, 0x80, 0x80, 0x80, "ds_add_src2_u32"),
1140    (0x81, 0x81, 0x81, 0x81, 0x81, "ds_sub_src2_u32"),
1141    (0x82, 0x82, 0x82, 0x82, 0x82, "ds_rsub_src2_u32"),
1142    (0x83, 0x83, 0x83, 0x83, 0x83, "ds_inc_src2_u32"),
1143    (0x84, 0x84, 0x84, 0x84, 0x84, "ds_dec_src2_u32"),
1144    (0x85, 0x85, 0x85, 0x85, 0x85, "ds_min_src2_i32"),
1145    (0x86, 0x86, 0x86, 0x86, 0x86, "ds_max_src2_i32"),
1146    (0x87, 0x87, 0x87, 0x87, 0x87, "ds_min_src2_u32"),
1147    (0x88, 0x88, 0x88, 0x88, 0x88, "ds_max_src2_u32"),
1148    (0x89, 0x89, 0x89, 0x89, 0x89, "ds_and_src2_b32"),
1149    (0x8a, 0x8a, 0x8a, 0x8a, 0x8a, "ds_or_src2_b32"),
1150    (0x8b, 0x8b, 0x8b, 0x8b, 0x8b, "ds_xor_src2_b32"),
1151    (0x8d, 0x8d, 0x8d, 0x8d, 0x8d, "ds_write_src2_b32"),
1152    (0x92, 0x92, 0x92, 0x92, 0x92, "ds_min_src2_f32"),
1153    (0x93, 0x93, 0x93, 0x93, 0x93, "ds_max_src2_f32"),
1154    (  -1,   -1, 0x95, 0x95, 0x95, "ds_add_src2_f32"),
1155    (  -1, 0x18, 0x98, 0x98, 0x18, "ds_gws_sema_release_all"),
1156    (0x19, 0x19, 0x99, 0x99, 0x19, "ds_gws_init"),
1157    (0x1a, 0x1a, 0x9a, 0x9a, 0x1a, "ds_gws_sema_v"),
1158    (0x1b, 0x1b, 0x9b, 0x9b, 0x1b, "ds_gws_sema_br"),
1159    (0x1c, 0x1c, 0x9c, 0x9c, 0x1c, "ds_gws_sema_p"),
1160    (0x1d, 0x1d, 0x9d, 0x9d, 0x1d, "ds_gws_barrier"),
1161    (  -1,   -1, 0xb6, 0xb6, 0xb1, "ds_read_addtid_b32"),
1162    (0x3d, 0x3d, 0xbd, 0xbd, 0x3d, "ds_consume"),
1163    (0x3e, 0x3e, 0xbe, 0xbe, 0x3e, "ds_append"),
1164    (0x3f, 0x3f, 0xbf, 0xbf, 0x3f, "ds_ordered_count"),
1165    (0xc0, 0xc0, 0xc0, 0xc0, 0xc0, "ds_add_src2_u64"),
1166    (0xc1, 0xc1, 0xc1, 0xc1, 0xc1, "ds_sub_src2_u64"),
1167    (0xc2, 0xc2, 0xc2, 0xc2, 0xc2, "ds_rsub_src2_u64"),
1168    (0xc3, 0xc3, 0xc3, 0xc3, 0xc3, "ds_inc_src2_u64"),
1169    (0xc4, 0xc4, 0xc4, 0xc4, 0xc4, "ds_dec_src2_u64"),
1170    (0xc5, 0xc5, 0xc5, 0xc5, 0xc5, "ds_min_src2_i64"),
1171    (0xc6, 0xc6, 0xc6, 0xc6, 0xc6, "ds_max_src2_i64"),
1172    (0xc7, 0xc7, 0xc7, 0xc7, 0xc7, "ds_min_src2_u64"),
1173    (0xc8, 0xc8, 0xc8, 0xc8, 0xc8, "ds_max_src2_u64"),
1174    (0xc9, 0xc9, 0xc9, 0xc9, 0xc9, "ds_and_src2_b64"),
1175    (0xca, 0xca, 0xca, 0xca, 0xca, "ds_or_src2_b64"),
1176    (0xcb, 0xcb, 0xcb, 0xcb, 0xcb, "ds_xor_src2_b64"),
1177    (0xcd, 0xcd, 0xcd, 0xcd, 0xcd, "ds_write_src2_b64"),
1178    (0xd2, 0xd2, 0xd2, 0xd2, 0xd2, "ds_min_src2_f64"),
1179    (0xd3, 0xd3, 0xd3, 0xd3, 0xd3, "ds_max_src2_f64"),
1180    (  -1, 0xde, 0xde, 0xde, 0xde, "ds_write_b96"),
1181    (  -1, 0xdf, 0xdf, 0xdf, 0xdf, "ds_write_b128"),
1182    (  -1, 0xfd, 0xfd,   -1,   -1, "ds_condxchg32_rtn_b128"),
1183    (  -1, 0xfe, 0xfe, 0xfe, 0xfe, "ds_read_b96"),
1184    (  -1, 0xff, 0xff, 0xff, 0xff, "ds_read_b128"),
1185 }
1186 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in DS:
1187     opcode(name, gfx7, gfx9, gfx10, Format.DS)
1188
1189 # MUBUF instructions:
1190 MUBUF = {
1191    (0x00, 0x00, 0x00, 0x00, 0x00, "buffer_load_format_x"),
1192    (0x01, 0x01, 0x01, 0x01, 0x01, "buffer_load_format_xy"),
1193    (0x02, 0x02, 0x02, 0x02, 0x02, "buffer_load_format_xyz"),
1194    (0x03, 0x03, 0x03, 0x03, 0x03, "buffer_load_format_xyzw"),
1195    (0x04, 0x04, 0x04, 0x04, 0x04, "buffer_store_format_x"),
1196    (0x05, 0x05, 0x05, 0x05, 0x05, "buffer_store_format_xy"),
1197    (0x06, 0x06, 0x06, 0x06, 0x06, "buffer_store_format_xyz"),
1198    (0x07, 0x07, 0x07, 0x07, 0x07, "buffer_store_format_xyzw"),
1199    (  -1,   -1, 0x08, 0x08, 0x80, "buffer_load_format_d16_x"),
1200    (  -1,   -1, 0x09, 0x09, 0x81, "buffer_load_format_d16_xy"),
1201    (  -1,   -1, 0x0a, 0x0a, 0x82, "buffer_load_format_d16_xyz"),
1202    (  -1,   -1, 0x0b, 0x0b, 0x83, "buffer_load_format_d16_xyzw"),
1203    (  -1,   -1, 0x0c, 0x0c, 0x84, "buffer_store_format_d16_x"),
1204    (  -1,   -1, 0x0d, 0x0d, 0x85, "buffer_store_format_d16_xy"),
1205    (  -1,   -1, 0x0e, 0x0e, 0x86, "buffer_store_format_d16_xyz"),
1206    (  -1,   -1, 0x0f, 0x0f, 0x87, "buffer_store_format_d16_xyzw"),
1207    (0x08, 0x08, 0x10, 0x10, 0x08, "buffer_load_ubyte"),
1208    (0x09, 0x09, 0x11, 0x11, 0x09, "buffer_load_sbyte"),
1209    (0x0a, 0x0a, 0x12, 0x12, 0x0a, "buffer_load_ushort"),
1210    (0x0b, 0x0b, 0x13, 0x13, 0x0b, "buffer_load_sshort"),
1211    (0x0c, 0x0c, 0x14, 0x14, 0x0c, "buffer_load_dword"),
1212    (0x0d, 0x0d, 0x15, 0x15, 0x0d, "buffer_load_dwordx2"),
1213    (  -1, 0x0f, 0x16, 0x16, 0x0f, "buffer_load_dwordx3"),
1214    (0x0f, 0x0e, 0x17, 0x17, 0x0e, "buffer_load_dwordx4"),
1215    (0x18, 0x18, 0x18, 0x18, 0x18, "buffer_store_byte"),
1216    (  -1,   -1,   -1, 0x19, 0x19, "buffer_store_byte_d16_hi"),
1217    (0x1a, 0x1a, 0x1a, 0x1a, 0x1a, "buffer_store_short"),
1218    (  -1,   -1,   -1, 0x1b, 0x1b, "buffer_store_short_d16_hi"),
1219    (0x1c, 0x1c, 0x1c, 0x1c, 0x1c, "buffer_store_dword"),
1220    (0x1d, 0x1d, 0x1d, 0x1d, 0x1d, "buffer_store_dwordx2"),
1221    (  -1, 0x1f, 0x1e, 0x1e, 0x1f, "buffer_store_dwordx3"),
1222    (0x1e, 0x1e, 0x1f, 0x1f, 0x1e, "buffer_store_dwordx4"),
1223    (  -1,   -1,   -1, 0x20, 0x20, "buffer_load_ubyte_d16"),
1224    (  -1,   -1,   -1, 0x21, 0x21, "buffer_load_ubyte_d16_hi"),
1225    (  -1,   -1,   -1, 0x22, 0x22, "buffer_load_sbyte_d16"),
1226    (  -1,   -1,   -1, 0x23, 0x23, "buffer_load_sbyte_d16_hi"),
1227    (  -1,   -1,   -1, 0x24, 0x24, "buffer_load_short_d16"),
1228    (  -1,   -1,   -1, 0x25, 0x25, "buffer_load_short_d16_hi"),
1229    (  -1,   -1,   -1, 0x26, 0x26, "buffer_load_format_d16_hi_x"),
1230    (  -1,   -1,   -1, 0x27, 0x27, "buffer_store_format_d16_hi_x"),
1231    (  -1,   -1, 0x3d, 0x3d,   -1, "buffer_store_lds_dword"),
1232    (0x71, 0x71, 0x3e, 0x3e,   -1, "buffer_wbinvl1"),
1233    (0x70, 0x70, 0x3f, 0x3f,   -1, "buffer_wbinvl1_vol"),
1234    (0x30, 0x30, 0x40, 0x40, 0x30, "buffer_atomic_swap"),
1235    (0x31, 0x31, 0x41, 0x41, 0x31, "buffer_atomic_cmpswap"),
1236    (0x32, 0x32, 0x42, 0x42, 0x32, "buffer_atomic_add"),
1237    (0x33, 0x33, 0x43, 0x43, 0x33, "buffer_atomic_sub"),
1238    (0x34,   -1,   -1,   -1,   -1, "buffer_atomic_rsub"),
1239    (0x35, 0x35, 0x44, 0x44, 0x35, "buffer_atomic_smin"),
1240    (0x36, 0x36, 0x45, 0x45, 0x36, "buffer_atomic_umin"),
1241    (0x37, 0x37, 0x46, 0x46, 0x37, "buffer_atomic_smax"),
1242    (0x38, 0x38, 0x47, 0x47, 0x38, "buffer_atomic_umax"),
1243    (0x39, 0x39, 0x48, 0x48, 0x39, "buffer_atomic_and"),
1244    (0x3a, 0x3a, 0x49, 0x49, 0x3a, "buffer_atomic_or"),
1245    (0x3b, 0x3b, 0x4a, 0x4a, 0x3b, "buffer_atomic_xor"),
1246    (0x3c, 0x3c, 0x4b, 0x4b, 0x3c, "buffer_atomic_inc"),
1247    (0x3d, 0x3d, 0x4c, 0x4c, 0x3d, "buffer_atomic_dec"),
1248    (0x3e, 0x3e,   -1,   -1, 0x3e, "buffer_atomic_fcmpswap"),
1249    (0x3f, 0x3f,   -1,   -1, 0x3f, "buffer_atomic_fmin"),
1250    (0x40, 0x40,   -1,   -1, 0x40, "buffer_atomic_fmax"),
1251    (0x50, 0x50, 0x60, 0x60, 0x50, "buffer_atomic_swap_x2"),
1252    (0x51, 0x51, 0x61, 0x61, 0x51, "buffer_atomic_cmpswap_x2"),
1253    (0x52, 0x52, 0x62, 0x62, 0x52, "buffer_atomic_add_x2"),
1254    (0x53, 0x53, 0x63, 0x63, 0x53, "buffer_atomic_sub_x2"),
1255    (0x54,   -1,   -1,   -1,   -1, "buffer_atomic_rsub_x2"),
1256    (0x55, 0x55, 0x64, 0x64, 0x55, "buffer_atomic_smin_x2"),
1257    (0x56, 0x56, 0x65, 0x65, 0x56, "buffer_atomic_umin_x2"),
1258    (0x57, 0x57, 0x66, 0x66, 0x57, "buffer_atomic_smax_x2"),
1259    (0x58, 0x58, 0x67, 0x67, 0x58, "buffer_atomic_umax_x2"),
1260    (0x59, 0x59, 0x68, 0x68, 0x59, "buffer_atomic_and_x2"),
1261    (0x5a, 0x5a, 0x69, 0x69, 0x5a, "buffer_atomic_or_x2"),
1262    (0x5b, 0x5b, 0x6a, 0x6a, 0x5b, "buffer_atomic_xor_x2"),
1263    (0x5c, 0x5c, 0x6b, 0x6b, 0x5c, "buffer_atomic_inc_x2"),
1264    (0x5d, 0x5d, 0x6c, 0x6c, 0x5d, "buffer_atomic_dec_x2"),
1265    (0x5e, 0x5e,   -1,   -1, 0x5e, "buffer_atomic_fcmpswap_x2"),
1266    (0x5f, 0x5f,   -1,   -1, 0x5f, "buffer_atomic_fmin_x2"),
1267    (0x60, 0x60,   -1,   -1, 0x60, "buffer_atomic_fmax_x2"),
1268    (  -1,   -1,   -1,   -1, 0x71, "buffer_gl0_inv"),
1269    (  -1,   -1,   -1,   -1, 0x72, "buffer_gl1_inv"),
1270 }
1271 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in MUBUF:
1272     opcode(name, gfx7, gfx9, gfx10, Format.MUBUF, is_atomic = "atomic" in name)
1273
1274 MTBUF = {
1275    (0x00, 0x00, 0x00, 0x00, 0x00, "tbuffer_load_format_x"),
1276    (0x01, 0x01, 0x01, 0x01, 0x01, "tbuffer_load_format_xy"),
1277    (0x02, 0x02, 0x02, 0x02, 0x02, "tbuffer_load_format_xyz"),
1278    (0x03, 0x03, 0x03, 0x03, 0x03, "tbuffer_load_format_xyzw"),
1279    (0x04, 0x04, 0x04, 0x04, 0x04, "tbuffer_store_format_x"),
1280    (0x05, 0x05, 0x05, 0x05, 0x05, "tbuffer_store_format_xy"),
1281    (0x06, 0x06, 0x06, 0x06, 0x06, "tbuffer_store_format_xyz"),
1282    (0x07, 0x07, 0x07, 0x07, 0x07, "tbuffer_store_format_xyzw"),
1283    (  -1,   -1, 0x08, 0x08, 0x08, "tbuffer_load_format_d16_x"),
1284    (  -1,   -1, 0x09, 0x09, 0x09, "tbuffer_load_format_d16_xy"),
1285    (  -1,   -1, 0x0a, 0x0a, 0x0a, "tbuffer_load_format_d16_xyz"),
1286    (  -1,   -1, 0x0b, 0x0b, 0x0b, "tbuffer_load_format_d16_xyzw"),
1287    (  -1,   -1, 0x0c, 0x0c, 0x0c, "tbuffer_store_format_d16_x"),
1288    (  -1,   -1, 0x0d, 0x0d, 0x0d, "tbuffer_store_format_d16_xy"),
1289    (  -1,   -1, 0x0e, 0x0e, 0x0e, "tbuffer_store_format_d16_xyz"),
1290    (  -1,   -1, 0x0f, 0x0f, 0x0f, "tbuffer_store_format_d16_xyzw"),
1291 }
1292 for (gfx6, gfx7, gfx8, gfx9, gfx10, name) in MTBUF:
1293     opcode(name, gfx7, gfx9, gfx10, Format.MTBUF)
1294
1295
1296 IMAGE = {
1297    (0x00, "image_load"),
1298    (0x01, "image_load_mip"),
1299    (0x02, "image_load_pck"),
1300    (0x03, "image_load_pck_sgn"),
1301    (0x04, "image_load_mip_pck"),
1302    (0x05, "image_load_mip_pck_sgn"),
1303    (0x08, "image_store"),
1304    (0x09, "image_store_mip"),
1305    (0x0a, "image_store_pck"),
1306    (0x0b, "image_store_mip_pck"),
1307    (0x0e, "image_get_resinfo"),
1308    (0x60, "image_get_lod"),
1309 }
1310 # (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (code, code, code, code, code, name)
1311 for (code, name) in IMAGE:
1312    opcode(name, code, code, code, Format.MIMG)
1313
1314 IMAGE_ATOMIC = {
1315    (0x0f, 0x0f, 0x10, "image_atomic_swap"),
1316    (0x10, 0x10, 0x11, "image_atomic_cmpswap"),
1317    (0x11, 0x11, 0x12, "image_atomic_add"),
1318    (0x12, 0x12, 0x13, "image_atomic_sub"),
1319    (0x13,   -1,   -1, "image_atomic_rsub"),
1320    (0x14, 0x14, 0x14, "image_atomic_smin"),
1321    (0x15, 0x15, 0x15, "image_atomic_umin"),
1322    (0x16, 0x16, 0x16, "image_atomic_smax"),
1323    (0x17, 0x17, 0x17, "image_atomic_umax"),
1324    (0x18, 0x18, 0x18, "image_atomic_and"),
1325    (0x19, 0x19, 0x19, "image_atomic_or"),
1326    (0x1a, 0x1a, 0x1a, "image_atomic_xor"),
1327    (0x1b, 0x1b, 0x1b, "image_atomic_inc"),
1328    (0x1c, 0x1c, 0x1c, "image_atomic_dec"),
1329    (0x1d, 0x1d,   -1, "image_atomic_fcmpswap"),
1330    (0x1e, 0x1e,   -1, "image_atomic_fmin"),
1331    (0x1f, 0x1f,   -1, "image_atomic_fmax"),
1332 }
1333 # (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (gfx6, gfx7, gfx89, gfx89, ???, name)
1334 # gfx7 and gfx10 opcodes are the same here
1335 for (gfx6, gfx7, gfx89, name) in IMAGE_ATOMIC:
1336    opcode(name, gfx7, gfx89, gfx7, Format.MIMG, is_atomic = True)
1337
1338 IMAGE_SAMPLE = {
1339    (0x20, "image_sample"),
1340    (0x21, "image_sample_cl"),
1341    (0x22, "image_sample_d"),
1342    (0x23, "image_sample_d_cl"),
1343    (0x24, "image_sample_l"),
1344    (0x25, "image_sample_b"),
1345    (0x26, "image_sample_b_cl"),
1346    (0x27, "image_sample_lz"),
1347    (0x28, "image_sample_c"),
1348    (0x29, "image_sample_c_cl"),
1349    (0x2a, "image_sample_c_d"),
1350    (0x2b, "image_sample_c_d_cl"),
1351    (0x2c, "image_sample_c_l"),
1352    (0x2d, "image_sample_c_b"),
1353    (0x2e, "image_sample_c_b_cl"),
1354    (0x2f, "image_sample_c_lz"),
1355    (0x30, "image_sample_o"),
1356    (0x31, "image_sample_cl_o"),
1357    (0x32, "image_sample_d_o"),
1358    (0x33, "image_sample_d_cl_o"),
1359    (0x34, "image_sample_l_o"),
1360    (0x35, "image_sample_b_o"),
1361    (0x36, "image_sample_b_cl_o"),
1362    (0x37, "image_sample_lz_o"),
1363    (0x38, "image_sample_c_o"),
1364    (0x39, "image_sample_c_cl_o"),
1365    (0x3a, "image_sample_c_d_o"),
1366    (0x3b, "image_sample_c_d_cl_o"),
1367    (0x3c, "image_sample_c_l_o"),
1368    (0x3d, "image_sample_c_b_o"),
1369    (0x3e, "image_sample_c_b_cl_o"),
1370    (0x3f, "image_sample_c_lz_o"),
1371    (0x68, "image_sample_cd"),
1372    (0x69, "image_sample_cd_cl"),
1373    (0x6a, "image_sample_c_cd"),
1374    (0x6b, "image_sample_c_cd_cl"),
1375    (0x6c, "image_sample_cd_o"),
1376    (0x6d, "image_sample_cd_cl_o"),
1377    (0x6e, "image_sample_c_cd_o"),
1378    (0x6f, "image_sample_c_cd_cl_o"),
1379 }
1380 # (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (code, code, code, code, code, name)
1381 for (code, name) in IMAGE_SAMPLE:
1382    opcode(name, code, code, code, Format.MIMG)
1383
1384 IMAGE_GATHER4 = {
1385    (0x40, "image_gather4"),
1386    (0x41, "image_gather4_cl"),
1387    #(0x42, "image_gather4h"), VEGA only?
1388    (0x44, "image_gather4_l"), # following instructions have different opcodes according to ISA sheet.
1389    (0x45, "image_gather4_b"),
1390    (0x46, "image_gather4_b_cl"),
1391    (0x47, "image_gather4_lz"),
1392    (0x48, "image_gather4_c"),
1393    (0x49, "image_gather4_c_cl"), # previous instructions have different opcodes according to ISA sheet.
1394    #(0x4a, "image_gather4h_pck"), VEGA only?
1395    #(0x4b, "image_gather8h_pck"), VGEA only?
1396    (0x4c, "image_gather4_c_l"),
1397    (0x4d, "image_gather4_c_b"),
1398    (0x4e, "image_gather4_c_b_cl"),
1399    (0x4f, "image_gather4_c_lz"),
1400    (0x50, "image_gather4_o"),
1401    (0x51, "image_gather4_cl_o"),
1402    (0x54, "image_gather4_l_o"),
1403    (0x55, "image_gather4_b_o"),
1404    (0x56, "image_gather4_b_cl_o"),
1405    (0x57, "image_gather4_lz_o"),
1406    (0x58, "image_gather4_c_o"),
1407    (0x59, "image_gather4_c_cl_o"),
1408    (0x5c, "image_gather4_c_l_o"),
1409    (0x5d, "image_gather4_c_b_o"),
1410    (0x5e, "image_gather4_c_b_cl_o"),
1411    (0x5f, "image_gather4_c_lz_o"),
1412 }
1413 # (gfx6, gfx7, gfx8, gfx9, gfx10, name) = (code, code, code, code, code, name)
1414 for (code, name) in IMAGE_GATHER4:
1415    opcode(name, code, code, code, Format.MIMG)
1416
1417
1418 FLAT = {
1419    #GFX7, GFX8_9, GFX10
1420    (0x08, 0x10, 0x08, "flat_load_ubyte"),
1421    (0x09, 0x11, 0x09, "flat_load_sbyte"),
1422    (0x0a, 0x12, 0x0a, "flat_load_ushort"),
1423    (0x0b, 0x13, 0x0b, "flat_load_sshort"),
1424    (0x0c, 0x14, 0x0c, "flat_load_dword"),
1425    (0x0d, 0x15, 0x0d, "flat_load_dwordx2"),
1426    (0x0f, 0x16, 0x0f, "flat_load_dwordx3"),
1427    (0x0e, 0x17, 0x0e, "flat_load_dwordx4"),
1428    (0x18, 0x18, 0x18, "flat_store_byte"),
1429    (  -1, 0x19, 0x19, "flat_store_byte_d16_hi"),
1430    (0x1a, 0x1a, 0x1a, "flat_store_short"),
1431    (  -1, 0x1b, 0x1b, "flat_store_short_d16_hi"),
1432    (0x1c, 0x1c, 0x1c, "flat_store_dword"),
1433    (0x1d, 0x1d, 0x1d, "flat_store_dwordx2"),
1434    (0x1f, 0x1e, 0x1f, "flat_store_dwordx3"),
1435    (0x1e, 0x1f, 0x1e, "flat_store_dwordx4"),
1436    (  -1, 0x20, 0x20, "flat_load_ubyte_d16"),
1437    (  -1, 0x21, 0x21, "flat_load_ubyte_d16_hi"),
1438    (  -1, 0x22, 0x22, "flat_load_sbyte_d16"),
1439    (  -1, 0x23, 0x23, "flat_load_sbyte_d16_hi"),
1440    (  -1, 0x24, 0x24, "flat_load_short_d16"),
1441    (  -1, 0x25, 0x25, "flat_load_short_d16_hi"),
1442    (0x30, 0x40, 0x30, "flat_atomic_swap"),
1443    (0x31, 0x41, 0x31, "flat_atomic_cmpswap"),
1444    (0x32, 0x42, 0x32, "flat_atomic_add"),
1445    (0x33, 0x43, 0x33, "flat_atomic_sub"),
1446    (0x35, 0x44, 0x35, "flat_atomic_smin"),
1447    (0x36, 0x45, 0x36, "flat_atomic_umin"),
1448    (0x37, 0x46, 0x37, "flat_atomic_smax"),
1449    (0x38, 0x47, 0x38, "flat_atomic_umax"),
1450    (0x39, 0x48, 0x39, "flat_atomic_and"),
1451    (0x3a, 0x49, 0x3a, "flat_atomic_or"),
1452    (0x3b, 0x4a, 0x3b, "flat_atomic_xor"),
1453    (0x3c, 0x4b, 0x3c, "flat_atomic_inc"),
1454    (0x3d, 0x4c, 0x3d, "flat_atomic_dec"),
1455    (0x3e,   -1, 0x3e, "flat_atomic_fcmpswap"),
1456    (0x3f,   -1, 0x3f, "flat_atomic_fmin"),
1457    (0x40,   -1, 0x40, "flat_atomic_fmax"),
1458    (0x50, 0x60, 0x50, "flat_atomic_swap_x2"),
1459    (0x51, 0x61, 0x51, "flat_atomic_cmpswap_x2"),
1460    (0x52, 0x62, 0x52, "flat_atomic_add_x2"),
1461    (0x53, 0x63, 0x53, "flat_atomic_sub_x2"),
1462    (0x55, 0x64, 0x55, "flat_atomic_smin_x2"),
1463    (0x56, 0x65, 0x56, "flat_atomic_umin_x2"),
1464    (0x57, 0x66, 0x57, "flat_atomic_smax_x2"),
1465    (0x58, 0x67, 0x58, "flat_atomic_umax_x2"),
1466    (0x59, 0x68, 0x59, "flat_atomic_and_x2"),
1467    (0x5a, 0x69, 0x5a, "flat_atomic_or_x2"),
1468    (0x5b, 0x6a, 0x5b, "flat_atomic_xor_x2"),
1469    (0x5c, 0x6b, 0x5c, "flat_atomic_inc_x2"),
1470    (0x5d, 0x6c, 0x5d, "flat_atomic_dec_x2"),
1471    (0x5e,   -1, 0x5e, "flat_atomic_fcmpswap_x2"),
1472    (0x5f,   -1, 0x5f, "flat_atomic_fmin_x2"),
1473    (0x60,   -1, 0x60, "flat_atomic_fmax_x2"),
1474 }
1475 for (gfx7, gfx8, gfx10, name) in FLAT:
1476     opcode(name, gfx7, gfx8, gfx10, Format.FLAT, is_atomic = "atomic" in name)
1477
1478 GLOBAL = {
1479    #GFX8_9, GFX10
1480    (0x10, 0x08, "global_load_ubyte"),
1481    (0x11, 0x09, "global_load_sbyte"),
1482    (0x12, 0x0a, "global_load_ushort"),
1483    (0x13, 0x0b, "global_load_sshort"),
1484    (0x14, 0x0c, "global_load_dword"),
1485    (0x15, 0x0d, "global_load_dwordx2"),
1486    (0x16, 0x0f, "global_load_dwordx3"),
1487    (0x17, 0x0e, "global_load_dwordx4"),
1488    (0x18, 0x18, "global_store_byte"),
1489    (0x19, 0x19, "global_store_byte_d16_hi"),
1490    (0x1a, 0x1a, "global_store_short"),
1491    (0x1b, 0x1b, "global_store_short_d16_hi"),
1492    (0x1c, 0x1c, "global_store_dword"),
1493    (0x1d, 0x1d, "global_store_dwordx2"),
1494    (0x1e, 0x1f, "global_store_dwordx3"),
1495    (0x1f, 0x1e, "global_store_dwordx4"),
1496    (0x20, 0x20, "global_load_ubyte_d16"),
1497    (0x21, 0x21, "global_load_ubyte_d16_hi"),
1498    (0x22, 0x22, "global_load_sbyte_d16"),
1499    (0x23, 0x23, "global_load_sbyte_d16_hi"),
1500    (0x24, 0x24, "global_load_short_d16"),
1501    (0x25, 0x25, "global_load_short_d16_hi"),
1502    (0x40, 0x30, "global_atomic_swap"),
1503    (0x41, 0x31, "global_atomic_cmpswap"),
1504    (0x42, 0x32, "global_atomic_add"),
1505    (0x43, 0x33, "global_atomic_sub"),
1506    (0x44, 0x35, "global_atomic_smin"),
1507    (0x45, 0x36, "global_atomic_umin"),
1508    (0x46, 0x37, "global_atomic_smax"),
1509    (0x47, 0x38, "global_atomic_umax"),
1510    (0x48, 0x39, "global_atomic_and"),
1511    (0x49, 0x3a, "global_atomic_or"),
1512    (0x4a, 0x3b, "global_atomic_xor"),
1513    (0x4b, 0x3c, "global_atomic_inc"),
1514    (0x4c, 0x3d, "global_atomic_dec"),
1515    (  -1, 0x3e, "global_atomic_fcmpswap"),
1516    (  -1, 0x3f, "global_atomic_fmin"),
1517    (  -1, 0x40, "global_atomic_fmax"),
1518    (0x60, 0x50, "global_atomic_swap_x2"),
1519    (0x61, 0x51, "global_atomic_cmpswap_x2"),
1520    (0x62, 0x52, "global_atomic_add_x2"),
1521    (0x63, 0x53, "global_atomic_sub_x2"),
1522    (0x64, 0x55, "global_atomic_smin_x2"),
1523    (0x65, 0x56, "global_atomic_umin_x2"),
1524    (0x66, 0x57, "global_atomic_smax_x2"),
1525    (0x67, 0x58, "global_atomic_umax_x2"),
1526    (0x68, 0x59, "global_atomic_and_x2"),
1527    (0x69, 0x5a, "global_atomic_or_x2"),
1528    (0x6a, 0x5b, "global_atomic_xor_x2"),
1529    (0x6b, 0x5c, "global_atomic_inc_x2"),
1530    (0x6c, 0x5d, "global_atomic_dec_x2"),
1531    (  -1, 0x5e, "global_atomic_fcmpswap_x2"),
1532    (  -1, 0x5f, "global_atomic_fmin_x2"),
1533    (  -1, 0x60, "global_atomic_fmax_x2"),
1534 }
1535 for (gfx8, gfx10, name) in GLOBAL:
1536     opcode(name, -1, gfx8, gfx10, Format.GLOBAL, is_atomic = "atomic" in name)
1537
1538 SCRATCH = {
1539    #GFX8_9, GFX10
1540    (0x10, 0x08, "scratch_load_ubyte"),
1541    (0x11, 0x09, "scratch_load_sbyte"),
1542    (0x12, 0x0a, "scratch_load_ushort"),
1543    (0x13, 0x0b, "scratch_load_sshort"),
1544    (0x14, 0x0c, "scratch_load_dword"),
1545    (0x15, 0x0d, "scratch_load_dwordx2"),
1546    (0x16, 0x0f, "scratch_load_dwordx3"),
1547    (0x17, 0x0e, "scratch_load_dwordx4"),
1548    (0x18, 0x18, "scratch_store_byte"),
1549    (0x19, 0x19, "scratch_store_byte_d16_hi"),
1550    (0x1a, 0x1a, "scratch_store_short"),
1551    (0x1b, 0x1b, "scratch_store_short_d16_hi"),
1552    (0x1c, 0x1c, "scratch_store_dword"),
1553    (0x1d, 0x1d, "scratch_store_dwordx2"),
1554    (0x1e, 0x1f, "scratch_store_dwordx3"),
1555    (0x1f, 0x1e, "scratch_store_dwordx4"),
1556    (0x20, 0x20, "scratch_load_ubyte_d16"),
1557    (0x21, 0x21, "scratch_load_ubyte_d16_hi"),
1558    (0x22, 0x22, "scratch_load_sbyte_d16"),
1559    (0x23, 0x23, "scratch_load_sbyte_d16_hi"),
1560    (0x24, 0x24, "scratch_load_short_d16"),
1561    (0x25, 0x25, "scratch_load_short_d16_hi"),
1562 }
1563 for (gfx8, gfx10, name) in SCRATCH:
1564     opcode(name, -1, gfx8, gfx10, Format.SCRATCH)
1565
1566 # check for duplicate opcode numbers
1567 for ver in ['gfx9', 'gfx10']:
1568     op_to_name = {}
1569     for op in opcodes.values():
1570         if op.format in [Format.PSEUDO, Format.PSEUDO_BRANCH, Format.PSEUDO_BARRIER, Format.PSEUDO_REDUCTION]:
1571             continue
1572
1573         num = getattr(op, 'opcode_' + ver)
1574         if num == -1:
1575             continue
1576
1577         key = (op.format, num)
1578
1579         if key in op_to_name:
1580             # exceptions
1581             names = set([op_to_name[key], op.name])
1582             if ver in ['gfx8', 'gfx9'] and names == set(['v_mul_lo_i32', 'v_mul_lo_u32']):
1583                 continue
1584
1585             print('%s and %s share the same opcode number (%s)' % (op_to_name[key], op.name, ver))
1586             sys.exit(1)
1587         else:
1588             op_to_name[key] = op.name