src/soc/sv/trans/svp64.py

   1 # SPDX-License-Identifier: LGPLv3+
   2 # Copyright (C) 2021 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
   3 # Funded by NLnet http://nlnet.nl
   4
   5 """SVP64 OpenPOWER v3.0B assembly translator
   6
   7 This class takes raw svp64 assembly mnemonics (aliases excluded) and
   8 creates an EXT001-encoded "svp64 prefix" followed by a v3.0B opcode.
   9
  10 It is very simple and straightforward, the only weirdness being the
  11 extraction of the register information and conversion to v3.0B numbering.
  12
  13 Encoding format of svp64: https://libre-soc.org/openpower/sv/svp64/
  14 Bugtracker: https://bugs.libre-soc.org/show_bug.cgi?id=578
  15 """
  16
  17 import os, sys
  18 from collections import OrderedDict
  19
  20 from soc.decoder.pseudo.pagereader import ISA
  21 from soc.decoder.power_enums import get_csv, find_wiki_dir
  22
  23
  24 # identifies register by type
  25 def is_CR_3bit(regname):
  26     return regname in ['BF', 'BFA']
  27
  28 def is_CR_5bit(regname):
  29     return regname in ['BA', 'BB', 'BC', 'BI', 'BT']
  30
  31 def is_GPR(regname):
  32     return regname in ['RA', 'RB', 'RC', 'RS', 'RT']
  33
  34 def get_regtype(regname):
  35     if is_CR_3bit(regname):
  36         return "CR_3bit"
  37     if is_CR_5bit(regname):
  38         return "CR_5bit"
  39     if is_GPR(regname):
  40         return "GPR"
  41
  42 # decode GPR into sv extra
  43 def  get_extra_gpr(etype, regmode, field):
  44     if regmode == 'scalar':
  45         # cut into 2-bits 5-bits SS FFFFF
  46         sv_extra = field >> 5
  47         field = field & 0b11111
  48     else:
  49         # cut into 5-bits 2-bits FFFFF SS
  50         sv_extra = field & 0b11
  51         field = field >> 2
  52     return sv_extra, field
  53
  54
  55 # decode 3-bit CR into sv extra
  56 def  get_extra_cr_3bit(etype, regmode, field):
  57     if regmode == 'scalar':
  58         # cut into 2-bits 3-bits SS FFF
  59         sv_extra = field >> 3
  60         field = field & 0b111
  61     else:
  62         # cut into 3-bits 4-bits FFF SSSS but will cut 2 zeros off later
  63         sv_extra = field & 0b1111
  64         field = field >> 4
  65     return sv_extra, field
  66
  67
  68 # decodes SUBVL
  69 def decode_subvl(encoding):
  70     pmap = {'2': 0b01, '3': 0b10, '4': 0b11}
  71     assert encoding in pmap, \
  72         "encoding %s for SUBVL not recognised" % encoding
  73     return pmap[encoding]
  74
  75
  76 # decodes elwidth
  77 def decode_elwidth(encoding):
  78     pmap = {'8': 0b11, '16': 0b10, '32': 0b01}
  79     assert encoding in pmap, \
  80         "encoding %s for elwidth not recognised" % encoding
  81     return pmap[encoding]
  82
  83
  84 # decodes predicate register encoding
  85 def decode_predicate(encoding):
  86     pmap = { # integer
  87             '1<<r3': (0, 0b001),
  88             'r3'   : (0, 0b010),
  89             '~r3'   : (0, 0b011),
  90             'r10'   : (0, 0b100),
  91             '~r10'  : (0, 0b101),
  92             'r30'   : (0, 0b110),
  93             '~r30'  : (0, 0b111),
  94             # CR
  95             'lt'    : (1, 0b000),
  96             'nl'    : (1, 0b001), 'ge'    : (1, 0b001), # same value
  97             'gt'    : (1, 0b010),
  98             'ng'    : (1, 0b011), 'le'    : (1, 0b011), # same value
  99             'eq'    : (1, 0b100),
 100             'ne'    : (1, 0b101),
 101             'so'    : (1, 0b110), 'un'    : (1, 0b110), # same value
 102             'ns'    : (1, 0b111), 'nu'    : (1, 0b111), # same value
 103            }
 104     assert encoding in pmap, \
 105         "encoding %s for predicate not recognised" % encoding
 106     return pmap[encoding]
 107
 108
 109 # decodes "Mode" in similar way to BO field (supposed to, anyway)
 110 def decode_bo(encoding):
 111     pmap = { # TODO: double-check that these are the same as Branch BO
 112             'lt'    : 0b000,
 113             'nl'    : 0b001, 'ge'    : 0b001, # same value
 114             'gt'    : 0b010,
 115             'ng'    : 0b011, 'le'    : 0b011, # same value
 116             'eq'    : 0b100,
 117             'ne'    : 0b101,
 118             'so'    : 0b110, 'un'    : 0b110, # same value
 119             'ns'    : 0b111, 'nu'    : 0b111, # same value
 120            }
 121     assert encoding in pmap, \
 122         "encoding %s for BO Mode not recognised" % encoding
 123     return pmap[encoding]
 124
 125 # partial-decode fail-first mode
 126 def decode_ffirst(encoding):
 127     if encoding in ['RC1', '~RC1']:
 128         return encoding
 129     return decode_bo(encoding)
 130
 131
 132 # gets SVP64 ReMap information
 133 class SVP64RM:
 134     def __init__(self):
 135         self.instrs = {}
 136         pth = find_wiki_dir()
 137         for fname in os.listdir(pth):
 138             if fname.startswith("RM"):
 139                 for entry in get_csv(fname):
 140                     self.instrs[entry['insn']] = entry
 141
 142
 143 # decodes svp64 assembly listings and creates EXT001 svp64 prefixes
 144 class SVP64:
 145     def __init__(self, lst):
 146         self.lst = lst
 147         self.trans = self.translate(lst)
 148
 149     def __iter__(self):
 150         for insn in self.trans:
 151             yield insn
 152
 153     def translate(self, lst):
 154         isa = ISA() # reads the v3.0B pseudo-code markdown files
 155         svp64 = SVP64RM() # reads the svp64 Remap entries for registers
 156         res = []
 157         for insn in lst:
 158             # find first space, to get opcode
 159             ls = insn.split(' ')
 160             opcode = ls[0]
 161             # now find opcode fields
 162             fields = ''.join(ls[1:]).split(',')
 163             fields = list(map(str.strip, fields))
 164             print ("opcode, fields", ls, opcode, fields)
 165
 166             # identify if is a svp64 mnemonic
 167             if not opcode.startswith('sv.'):
 168                 res.append(insn) # unaltered
 169                 continue
 170             opcode = opcode[3:] # strip leading "sv."
 171
 172             # start working on decoding the svp64 op: sv.baseop/vec2.mode
 173             opcode = opcode.split("/") # split at "/"
 174             v30b_op = opcode[0]       # first is the v3.0B
 175             if len(opcode) == 1:
 176                 opmodes = [] # no sv modes
 177             else:
 178                 opmodes = opcode[1].split(".") # second splits by dots
 179
 180             # check instruction ends with dot
 181             rc_mode = v30b_op.endswith('.')
 182             if rc_mode:
 183                 v30b_op = v30b_op[:-1]
 184
 185             if v30b_op not in isa.instr:
 186                 raise Exception("opcode %s of '%s' not supported" % \
 187                                 (v30b_op, insn))
 188             if v30b_op not in svp64.instrs:
 189                 raise Exception("opcode %s of '%s' not an svp64 instruction" % \
 190                                 (v30b_op, insn))
 191             isa.instr[v30b_op].regs[0]
 192             v30b_regs = isa.instr[v30b_op].regs[0]
 193             rm = svp64.instrs[v30b_op]
 194             print ("v3.0B op", v30b_op, "Rc=1" if rc_mode else '')
 195             print ("v3.0B regs", opcode, v30b_regs)
 196             print (rm)
 197
 198             # right.  the first thing to do is identify the ordering of
 199             # the registers, by name.  the EXTRA2/3 ordering is in
 200             # rm['0']..rm['3'] but those fields contain the names RA, BB
 201             # etc.  we have to read the pseudocode to understand which
 202             # reg is which in our instruction. sigh.
 203
 204             # first turn the svp64 rm into a "by name" dict, recording
 205             # which position in the RM EXTRA it goes into
 206             # also: record if the src or dest was a CR, for sanity-checking
 207             # (elwidth overrides on CRs are banned)
 208             dest_reg_cr, src_reg_cr = False, False
 209             svp64_reg_byname = {}
 210             for i in range(4):
 211                 rfield = rm[str(i)]
 212                 if not rfield or rfield == '0':
 213                     continue
 214                 print ("EXTRA field", i, rfield)
 215                 rfield = rfield.split(";") # s:RA;d:CR1 etc.
 216                 for r in rfield:
 217                     rtype = r[0]
 218                     r = r[2:] # ignore s: and d:
 219                     svp64_reg_byname[r] = i # this reg in EXTRA position 0-3
 220                     # check the regtype (if CR, record that)
 221                     regtype = get_regtype(r)
 222                     if regtype in ['CR_3bit', 'CR_5bit']:
 223                         if rtype == 'd':
 224                             dest_reg_cr = True
 225                         if rtype == 'd':
 226                             src_reg_cr = True
 227
 228             print ("EXTRA field index, by regname", svp64_reg_byname)
 229
 230             # okaaay now we identify the field value (opcode N,N,N) with
 231             # the pseudo-code info (opcode RT, RA, RB)
 232             opregfields = zip(fields, v30b_regs) # err that was easy
 233
 234             # now for each of those find its place in the EXTRA encoding
 235             extras = OrderedDict()
 236             for idx, (field, regname) in enumerate(opregfields):
 237                 extra = svp64_reg_byname.get(regname, None)
 238                 regtype = get_regtype(regname)
 239                 extras[extra] = (idx, field, regname, regtype)
 240                 print ("    ", extra, extras[extra])
 241
 242             # great! got the extra fields in their associated positions:
 243             # also we know the register type. now to create the EXTRA encodings
 244             etype = rm['Etype'] # Extra type: EXTRA3/EXTRA2
 245             ptype = rm['Ptype'] # Predication type: Twin / Single
 246             extra_bits = 0
 247             v30b_newfields = []
 248             for extra_idx, (idx, field, regname, regtype) in extras.items():
 249                 # is it a field we don't alter/examine?  if so just put it
 250                 # into newfields
 251                 if regtype is None:
 252                     v30b_newfields.append(field)
 253
 254                 # first, decode the field number. "5.v" or "3.s" or "9"
 255                 field = field.split(".")
 256                 regmode = 'scalar' # default
 257                 if len(field) == 2:
 258                     if field[1] == 's':
 259                         regmode = 'scalar'
 260                     elif field[1] == 'v':
 261                         regmode = 'vector'
 262                 field = int(field[0]) # actual register number
 263                 print ("    ", regmode, field, end=" ")
 264
 265                 # XXX TODO: the following is a bit of a laborious repeated
 266                 # mess, which could (and should) easily be parameterised.
 267
 268                 # encode SV-GPR field into extra, v3.0field
 269                 if regtype == 'GPR':
 270                     sv_extra, field = get_extra_gpr(etype, regmode, field)
 271                     # now sanity-check. EXTRA3 is ok, EXTRA2 has limits
 272                     # (and shrink to a single bit if ok)
 273                     if etype == 'EXTRA2':
 274                         if regmode == 'scalar':
 275                             # range is r0-r63 in increments of 1
 276                             assert (sv_extra >> 1) == 0, \
 277                                 "scalar GPR %s cannot fit into EXTRA2 %s" % \
 278                                     (regname, str(extras[extra_idx]))
 279                             # all good: encode as scalar
 280                             sv_extra = sv_extra & 0b01
 281                         else:
 282                             # range is r0-r127 in increments of 4
 283                             assert sv_extra & 0b01 == 0, \
 284                                 "vector field %s cannot fit into EXTRA2 %s" % \
 285                                     (regname, str(extras[extra_idx]))
 286                             # all good: encode as vector (bit 2 set)
 287                             sv_extra = 0b10 | (sv_extra >> 1)
 288                     elif regmode == 'vector':
 289                         # EXTRA3 vector bit needs marking
 290                         sv_extra |= 0b100
 291
 292                 # encode SV-CR 3-bit field into extra, v3.0field
 293                 elif regtype == 'CR_3bit':
 294                     sv_extra, field = get_extra_cr_3bit(etype, regmode, field)
 295                     # now sanity-check (and shrink afterwards)
 296                     if etype == 'EXTRA2':
 297                         if regmode == 'scalar':
 298                             # range is CR0-CR15 in increments of 1
 299                             assert (sv_extra >> 1) == 0, \
 300                                 "scalar CR %s cannot fit into EXTRA2 %s" % \
 301                                     (regname, str(extras[extra_idx]))
 302                             # all good: encode as scalar
 303                             sv_extra = sv_extra & 0b01
 304                         else:
 305                             # range is CR0-CR127 in increments of 16
 306                             assert sv_extra & 0b111 == 0, \
 307                                 "vector CR %s cannot fit into EXTRA2 %s" % \
 308                                     (regname, str(extras[extra_idx]))
 309                             # all good: encode as vector (bit 2 set)
 310                             sv_extra = 0b10 | (sv_extra >> 3)
 311                     else:
 312                         if regmode == 'scalar':
 313                             # range is CR0-CR31 in increments of 1
 314                             assert (sv_extra >> 2) == 0, \
 315                                 "scalar CR %s cannot fit into EXTRA2 %s" % \
 316                                     (regname, str(extras[extra_idx]))
 317                             # all good: encode as scalar
 318                             sv_extra = sv_extra & 0b11
 319                         else:
 320                             # range is CR0-CR127 in increments of 8
 321                             assert sv_extra & 0b11 == 0, \
 322                                 "vector CR %s cannot fit into EXTRA2 %s" % \
 323                                     (regname, str(extras[extra_idx]))
 324                             # all good: encode as vector (bit 3 set)
 325                             sv_extra = 0b100 | (sv_extra >> 2)
 326
 327                 # encode SV-CR 5-bit field into extra, v3.0field
 328                 # *sigh* this is the same as 3-bit except the 2 LSBs are
 329                 # passed through
 330                 elif regtype == 'CR_5bit':
 331                     cr_subfield = field & 0b11
 332                     field = field >> 2 # strip bottom 2 bits
 333                     sv_extra, field = get_extra_cr_3bit(etype, regmode, field)
 334                     # now sanity-check (and shrink afterwards)
 335                     if etype == 'EXTRA2':
 336                         if regmode == 'scalar':
 337                             # range is CR0-CR15 in increments of 1
 338                             assert (sv_extra >> 1) == 0, \
 339                                 "scalar CR %s cannot fit into EXTRA2 %s" % \
 340                                     (regname, str(extras[extra_idx]))
 341                             # all good: encode as scalar
 342                             sv_extra = sv_extra & 0b01
 343                         else:
 344                             # range is CR0-CR127 in increments of 16
 345                             assert sv_extra & 0b111 == 0, \
 346                                 "vector CR %s cannot fit into EXTRA2 %s" % \
 347                                     (regname, str(extras[extra_idx]))
 348                             # all good: encode as vector (bit 2 set)
 349                             sv_extra = 0b10 | (sv_extra >> 3)
 350                     else:
 351                         if regmode == 'scalar':
 352                             # range is CR0-CR31 in increments of 1
 353                             assert (sv_extra >> 2) == 0, \
 354                                 "scalar CR %s cannot fit into EXTRA2 %s" % \
 355                                     (regname, str(extras[extra_idx]))
 356                             # all good: encode as scalar
 357                             sv_extra = sv_extra & 0b11
 358                         else:
 359                             # range is CR0-CR127 in increments of 8
 360                             assert sv_extra & 0b11 == 0, \
 361                                 "vector CR %s cannot fit into EXTRA2 %s" % \
 362                                     (regname, str(extras[extra_idx]))
 363                             # all good: encode as vector (bit 3 set)
 364                             sv_extra = 0b100 | (sv_extra >> 2)
 365
 366                     # reconstruct the actual 5-bit CR field
 367                     field = (field << 2) | cr_subfield
 368
 369                 # capture the extra field info
 370                 print ("=>", "%5s" % bin(sv_extra), field)
 371                 extras[extra_idx] = sv_extra
 372
 373                 # append altered field value to v3.0b
 374                 v30b_newfields.append(str(field))
 375
 376             print ("new v3.0B fields", v30b_op, v30b_newfields)
 377             print ("extras", extras)
 378
 379             # rright.  now we have all the info. start creating SVP64 RM
 380             svp64_rm = 0b0
 381
 382             # begin with EXTRA fields
 383             for idx, sv_extra in extras.items():
 384                 if idx is None: continue
 385                 # start at bit 10, work up 2/3 times EXTRA idx
 386                 offs = 2 if etype == 'EXTRA2' else 3 # 2 or 3 bits
 387                 svp64_rm |= sv_extra << (10+idx*offs)
 388
 389             # parts of svp64_rm
 390             mmode = 0  # bit 0
 391             pmask = 0  # bits 1-3
 392             destwid = 0 # bits 4-5
 393             srcwid = 0 # bits 6-7
 394             subvl = 0   # bits 8-9
 395             smask = 0 # bits 16-18 but only for twin-predication
 396             mode = 0 # bits 19-23
 397
 398             has_pmask = False
 399             has_smask = False
 400
 401             saturation = None
 402             src_zero = 0
 403             dst_zero = 0
 404             sv_mode = None
 405
 406             mapreduce = False
 407             mapreduce_crm = False
 408             mapreduce_svm = False
 409
 410             predresult = False
 411             failfirst = False
 412
 413             # ok let's start identifying opcode augmentation fields
 414             for encmode in opmodes:
 415                 # predicate mask (dest)
 416                 if encmode.startswith("m="):
 417                     pme = encmode
 418                     pmmode, pmask = decode_predicate(encmode[2:])
 419                     mmode = pmmode
 420                     has_pmask = True
 421                 # predicate mask (src, twin-pred)
 422                 elif encmode.startswith("sm="):
 423                     sme = encmode
 424                     smmode, smask = decode_predicate(encmode[3:])
 425                     mmode = smmode
 426                     has_smask = True
 427                 # vec2/3/4
 428                 elif encmode.startswith("vec"):
 429                     subvl = decode_subvl(encmode[3:])
 430                 # elwidth
 431                 elif encmode.startswith("ew="):
 432                     destwid = decode_elwidth(encmode[3:])
 433                 elif encmode.startswith("sw="):
 434                     srcwid = decode_elwidth(encmode[3:])
 435                 # saturation
 436                 elif encmode == 'sats':
 437                     assert sv_mode is None
 438                     saturation = 1
 439                     sv_mode = 0b10
 440                 elif encmode == 'satu':
 441                     assert sv_mode is None
 442                     sv_mode = 0b10
 443                     saturation = 0
 444                 # predicate zeroing
 445                 elif encmode == 'sz':
 446                     src_zero = 1
 447                 elif encmode == 'dz':
 448                     dst_zero = 1
 449                 # failfirst
 450                 elif encmode.startswith("ff="):
 451                     assert sv_mode is None
 452                     sv_mode = 0b01
 453                     failfirst = decode_ffirst(encmode[3:])
 454                 # predicate-result, interestingly same as fail-first
 455                 elif encmode.startswith("pr="):
 456                     assert sv_mode is None
 457                     sv_mode = 0b11
 458                     predresult = decode_ffirst(encmode[3:])
 459                 # map-reduce mode
 460                 elif encmode == 'mr':
 461                     assert sv_mode is None
 462                     sv_mode = 0b00
 463                     mapreduce = True
 464                 elif encmode == 'crm': # CR on map-reduce
 465                     assert sv_mode is None
 466                     sv_mode = 0b00
 467                     mapreduce_crm = True
 468                 elif encmode == 'svm': # sub-vector mode
 469                     mapreduce_svm = True
 470
 471             # construct the mode field, doing sanity-checking along the way
 472
 473             if mapreduce_svm:
 474                 assert sv_mode == 0b00, "sub-vector mode in mapreduce only"
 475                 assert subvl != 0, "sub-vector mode not possible on SUBVL=1"
 476
 477             # "normal" mode
 478             if sv_mode is None:
 479                 mode |= (src_zero << 3) | (dst_zero << 4)
 480
 481             # "mapreduce" modes
 482             elif sv_mode == 0b00:
 483                 mode |= (0b1<<2) # sets mapreduce
 484                 assert dst_zero == 0, "dest-zero not allowed in mapreduce mode"
 485                 if mapreduce_crm:
 486                     mode |= (0b1<<4) # sets CRM mode
 487                     assert rc_mode, "CRM only allowed when Rc=1"
 488                 # bit of weird encoding to jam zero-pred or SVM mode in.
 489                 # SVM mode can be enabled only when SUBVL=2/3/4 (vec2/3/4)
 490                 if subvl == 0:
 491                     mode |= (src_zero << 3) # predicate src-zeroing
 492                 elif mapreduce_svm:
 493                     mode |= (1 << 3) # SVM mode
 494
 495             # "failfirst" modes
 496             elif sv_mode == 0b01:
 497                 assert dst_zero == 0, "dest-zero not allowed in failfirst mode"
 498                 mode |= 0b01 # sets failfirst
 499                 if failfirst == 'RC1':
 500                     mode |= (0b1<<4) # sets RC1 mode
 501                     mode |= (src_zero << 3) # predicate src-zeroing
 502                     assert rc_mode==False, "ffirst RC1 only possible when Rc=0"
 503                 elif failfirst == '~RC1':
 504                     mode |= (0b1<<4) # sets RC1 mode...
 505                     mode |= (src_zero << 3) # predicate src-zeroing
 506                     mode |= (0b1<<2) # ... with inversion
 507                     assert rc_mode==False, "ffirst RC1 only possible when Rc=0"
 508                 else:
 509                     assert src_zero == 0, "src-zero not allowed in ffirst BO"
 510                     mode |= (failfirst << 2) # set BO
 511
 512             # sanity-check that 2Pred mask is same mode
 513             if has_pmask and has_smask:
 514                 assert smmode == pmmode, \
 515                     "predicate masks %s and %s must be same reg type" % \
 516                         (pme, sme)
 517
 518             # sanity-check that twin-predication mask only specified in 2P mode
 519             if ptype == '1P':
 520                 assert has_smask == False, \
 521                     "source-mask can only be specified on Twin-predicate ops"
 522
 523             # put in predicate masks into svp64_rm
 524             if ptype == '2P':
 525                 svp64_rm |= (smask << 16) # source pred: bits 16-18
 526             svp64_rm |= (mmode)           # mask mode: bit 0
 527             svp64_rm |= (pmask << 1)      # 1-pred: bits 1-3
 528
 529             # and subvl
 530             svp64_rm += (subvl << 8)      # subvl: bits 8-9
 531
 532             # put in elwidths
 533             svp64_rm += (srcwid << 6)      # srcwid: bits 6-7
 534             svp64_rm += (destwid << 4)     # destwid: bits 4-5
 535
 536             # nice debug printout. (and now for something completely different)
 537             # https://youtu.be/u0WOIwlXE9g?t=146
 538             print ("svp64_rm", hex(svp64_rm), bin(svp64_rm))
 539             print ("    mmode  0    :", bin(mmode))
 540             print ("    pmask  1-3  :", bin(pmask))
 541             print ("    dstwid 4-5  :", bin(destwid))
 542             print ("    srcwid 6-7  :", bin(srcwid))
 543             print ("    subvl  8-9  :", bin(subvl))
 544             print ("    mode   19-23:", bin(mode))
 545             offs = 2 if etype == 'EXTRA2' else 3 # 2 or 3 bits
 546             for idx, sv_extra in extras.items():
 547                 if idx is None: continue
 548                 start = (10+idx*offs)
 549                 end = start + offs-1
 550                 print ("    extra%d %2d-%2d:" % (idx, start, end),
 551                         bin(sv_extra))
 552             if ptype == '2P':
 553                 print ("    smask  16-17:", bin(smask))
 554             print ()
 555
 556         return res
 557
 558 if __name__ == '__main__':
 559     isa = SVP64(['slw 3, 1, 4',
 560                  'extsw 5, 3',
 561                  'sv.extsw 5, 3',
 562                  'sv.cmpi 5, 1, 3, 2',
 563                  'sv.setb 5, 31',
 564                  'sv.isel 64.v, 3, 2, 65.v',
 565                  'sv.setb/m=r3.sm=1<<r3 5, 31',
 566                  'sv.setb/vec2 5, 31',
 567                  'sv.setb/sw=8.ew=16 5, 31',
 568                  'sv.extsw./ff=eq 5, 31',
 569                 ])
 570     csvs = SVP64RM()