src/soc/decoder/isa/caller.py

   1 """core of the python-based POWER9 simulator
   2
   3 this is part of a cycle-accurate POWER9 simulator.  its primary purpose is
   4 not speed, it is for both learning and educational purposes, as well as
   5 a method of verifying the HDL.
   6 """
   7
   8 from functools import wraps
   9 from soc.decoder.orderedset import OrderedSet
  10 from soc.decoder.selectable_int import (FieldSelectableInt, SelectableInt,
  11                                         selectconcat)
  12 from soc.decoder.power_enums import spr_dict, XER_bits
  13 from soc.decoder.helpers import exts
  14 from collections import namedtuple
  15 import math
  16
  17 instruction_info = namedtuple('instruction_info',
  18                               'func read_regs uninit_regs write_regs ' + \
  19                               'special_regs op_fields form asmregs')
  20
  21 special_sprs = {
  22     'LR': 8,
  23     'CTR': 9,
  24     'TAR': 815,
  25     'XER': 1,
  26     'VRSAVE': 256}
  27
  28
  29 def swap_order(x, nbytes):
  30     x = x.to_bytes(nbytes, byteorder='little')
  31     x = int.from_bytes(x, byteorder='big', signed=False)
  32     return x
  33
  34
  35 def create_args(reglist, extra=None):
  36     args = OrderedSet()
  37     for reg in reglist:
  38         args.add(reg)
  39     args = list(args)
  40     if extra:
  41         args = [extra] + args
  42     return args
  43
  44
  45 class Mem:
  46
  47     def __init__(self, row_bytes=8, initial_mem=None):
  48         self.mem = {}
  49         self.bytes_per_word = row_bytes
  50         self.word_log2 = math.ceil(math.log2(row_bytes))
  51         print ("Sim-Mem", initial_mem, self.bytes_per_word, self.word_log2)
  52         if not initial_mem:
  53             return
  54
  55         # different types of memory data structures recognised (for convenience)
  56         if isinstance(initial_mem, list):
  57             initial_mem = (0, initial_mem)
  58         if isinstance(initial_mem, tuple):
  59             startaddr, mem = initial_mem
  60             initial_mem = {}
  61             for i, val in enumerate(mem):
  62                 initial_mem[startaddr + row_bytes*i] = (val, row_bytes)
  63
  64         for addr, (val, width) in initial_mem.items():
  65             #val = swap_order(val, width)
  66             self.st(addr, val, width, swap=False)
  67
  68     def _get_shifter_mask(self, wid, remainder):
  69         shifter = ((self.bytes_per_word - wid) - remainder) * \
  70             8  # bits per byte
  71         # XXX https://bugs.libre-soc.org/show_bug.cgi?id=377
  72         # BE/LE mode?
  73         shifter = remainder * 8
  74         mask = (1 << (wid * 8)) - 1
  75         print ("width,rem,shift,mask", wid, remainder, hex(shifter), hex(mask))
  76         return shifter, mask
  77
  78     # TODO: Implement ld/st of lesser width
  79     def ld(self, address, width=8, swap=True):
  80         print("ld from addr 0x{:x} width {:d}".format(address, width))
  81         remainder = address & (self.bytes_per_word - 1)
  82         address = address >> self.word_log2
  83         assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
  84         if address in self.mem:
  85             val = self.mem[address]
  86         else:
  87             val = 0
  88         print("mem @ 0x{:x} rem {:d} : 0x{:x}".format(address, remainder, val))
  89
  90         if width != self.bytes_per_word:
  91             shifter, mask = self._get_shifter_mask(width, remainder)
  92             print ("masking", hex(val), hex(mask<<shifter), shifter)
  93             val = val & (mask << shifter)
  94             val >>= shifter
  95         if swap:
  96             val = swap_order(val, width)
  97         print("Read 0x{:x} from addr 0x{:x}".format(val, address))
  98         return val
  99
 100     def st(self, addr, v, width=8, swap=True):
 101         staddr = addr
 102         remainder = addr & (self.bytes_per_word - 1)
 103         addr = addr >> self.word_log2
 104         print("Writing 0x{:x} to ST 0x{:x} memaddr 0x{:x}/{:x}".format(v,
 105                         staddr, addr, remainder, swap))
 106         assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
 107         if swap:
 108             v = swap_order(v, width)
 109         if width != self.bytes_per_word:
 110             if addr in self.mem:
 111                 val = self.mem[addr]
 112             else:
 113                 val = 0
 114             shifter, mask = self._get_shifter_mask(width, remainder)
 115             val &= ~(mask << shifter)
 116             val |= v << shifter
 117             self.mem[addr] = val
 118         else:
 119             self.mem[addr] = v
 120         print("mem @ 0x{:x}: 0x{:x}".format(addr, self.mem[addr]))
 121
 122     def __call__(self, addr, sz):
 123         val = self.ld(addr.value, sz)
 124         print ("memread", addr, sz, val)
 125         return SelectableInt(val, sz*8)
 126
 127     def memassign(self, addr, sz, val):
 128         print ("memassign", addr, sz, val)
 129         self.st(addr.value, val.value, sz)
 130
 131
 132 class GPR(dict):
 133     def __init__(self, decoder, regfile):
 134         dict.__init__(self)
 135         self.sd = decoder
 136         for i in range(32):
 137             self[i] = SelectableInt(regfile[i], 64)
 138
 139     def __call__(self, ridx):
 140         return self[ridx]
 141
 142     def set_form(self, form):
 143         self.form = form
 144
 145     def getz(self, rnum):
 146         #rnum = rnum.value # only SelectableInt allowed
 147         print("GPR getzero", rnum)
 148         if rnum == 0:
 149             return SelectableInt(0, 64)
 150         return self[rnum]
 151
 152     def _get_regnum(self, attr):
 153         getform = self.sd.sigforms[self.form]
 154         rnum = getattr(getform, attr)
 155         return rnum
 156
 157     def ___getitem__(self, attr):
 158         print("GPR getitem", attr)
 159         rnum = self._get_regnum(attr)
 160         return self.regfile[rnum]
 161
 162     def dump(self):
 163         for i in range(0, len(self), 8):
 164             s = []
 165             for j in range(8):
 166                 s.append("%08x" % self[i+j].value)
 167             s = ' '.join(s)
 168             print("reg", "%2d" % i, s)
 169
 170 class PC:
 171     def __init__(self, pc_init=0):
 172         self.CIA = SelectableInt(pc_init, 64)
 173         self.NIA = self.CIA + SelectableInt(4, 64)
 174
 175     def update(self, namespace):
 176         self.CIA = namespace['NIA'].narrow(64)
 177         self.NIA = self.CIA + SelectableInt(4, 64)
 178         namespace['CIA'] = self.CIA
 179         namespace['NIA'] = self.NIA
 180
 181
 182 class SPR(dict):
 183     def __init__(self, dec2, initial_sprs={}):
 184         self.sd = dec2
 185         dict.__init__(self)
 186         self.update(initial_sprs)
 187
 188     def __getitem__(self, key):
 189         # if key in special_sprs get the special spr, otherwise return key
 190         if isinstance(key, SelectableInt):
 191             key = key.value
 192         key = special_sprs.get(key, key)
 193         if key in self:
 194             return dict.__getitem__(self, key)
 195         else:
 196             info = spr_dict[key]
 197             dict.__setitem__(self, key, SelectableInt(0, info.length))
 198             return dict.__getitem__(self, key)
 199
 200     def __setitem__(self, key, value):
 201         if isinstance(key, SelectableInt):
 202             key = key.value
 203         key = special_sprs.get(key, key)
 204         dict.__setitem__(self, key, value)
 205
 206     def __call__(self, ridx):
 207         return self[ridx]
 208
 209
 210 class ISACaller:
 211     # decoder2 - an instance of power_decoder2
 212     # regfile - a list of initial values for the registers
 213     # initial_{etc} - initial values for SPRs, Condition Register, Mem, MSR
 214     # respect_pc - tracks the program counter.  requires initial_insns
 215     def __init__(self, decoder2, regfile, initial_sprs=None, initial_cr=0,
 216                        initial_mem=None, initial_msr=0,
 217                        initial_insns=None, respect_pc=False,
 218                        disassembly=None):
 219
 220         self.respect_pc = respect_pc
 221         if initial_sprs is None:
 222             initial_sprs = {}
 223         if initial_mem is None:
 224             initial_mem = {}
 225         if initial_insns is None:
 226             initial_insns = {}
 227             assert self.respect_pc == False, "instructions required to honor pc"
 228
 229
 230         # "fake program counter" mode (for unit testing)
 231         if not respect_pc:
 232             if isinstance(initial_mem, tuple):
 233                 self.fake_pc = initial_mem[0]
 234             else:
 235                 self.fake_pc = 0
 236
 237         # disassembly: we need this for now (not given from the decoder)
 238         self.disassembly = {}
 239         if disassembly:
 240             for i, code in enumerate(disassembly):
 241                 self.disassembly[i*4 + self.fake_pc] = code
 242
 243         self.gpr = GPR(decoder2, regfile)
 244         self.mem = Mem(row_bytes=8, initial_mem=initial_mem)
 245         self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
 246         self.pc = PC()
 247         self.spr = SPR(decoder2, initial_sprs)
 248         self.msr = SelectableInt(initial_msr, 64) # underlying reg
 249         # TODO, needed here:
 250         # FPR (same as GPR except for FP nums)
 251         # 4.2.2 p124 FPSCR (definitely "separate" - not in SPR)
 252         #            note that mffs, mcrfs, mtfsf "manage" this FPSCR
 253         # 2.3.1 CR (and sub-fields CR0..CR6 - CR0 SO comes from XER.SO)
 254         #         note that mfocrf, mfcr, mtcr, mtocrf, mcrxrx "manage" CRs
 255         #         -- Done
 256         # 2.3.2 LR   (actually SPR #8) -- Done
 257         # 2.3.3 CTR  (actually SPR #9) -- Done
 258         # 2.3.4 TAR  (actually SPR #815)
 259         # 3.2.2 p45 XER  (actually SPR #1) -- Done
 260         # 3.2.3 p46 p232 VRSAVE (actually SPR #256)
 261
 262         # create CR then allow portions of it to be "selectable" (below)
 263         self._cr = SelectableInt(initial_cr, 64) # underlying reg
 264         self.cr = FieldSelectableInt(self._cr, list(range(32,64)))
 265
 266         # "undefined", just set to variable-bit-width int (use exts "max")
 267         self.undefined = SelectableInt(0, 256) # TODO, not hard-code 256!
 268
 269         self.namespace = {'GPR': self.gpr,
 270                           'MEM': self.mem,
 271                           'SPR': self.spr,
 272                           'memassign': self.memassign,
 273                           'NIA': self.pc.NIA,
 274                           'CIA': self.pc.CIA,
 275                           'CR': self.cr,
 276                           'MSR': self.msr,
 277                           'undefined': self.undefined,
 278                           'mode_is_64bit': True,
 279                           'SO': XER_bits['SO']
 280                           }
 281
 282         # field-selectable versions of Condition Register TODO check bitranges?
 283         self.crl = []
 284         for i in range(8):
 285             bits = tuple(range(i*4, (i+1)*4))# errr... maybe?
 286             _cr = FieldSelectableInt(self.cr, bits)
 287             self.crl.append(_cr)
 288             self.namespace["CR%d" % i] = _cr
 289
 290         self.decoder = decoder2.dec
 291         self.dec2 = decoder2
 292
 293     def TRAP(self, trap_addr=0x700):
 294         print ("TRAP: TODO")
 295         # store CIA(+4?) in SRR0, set NIA to 0x700
 296         # store MSR in SRR1, set MSR to um errr something, have to check spec
 297
 298     def memassign(self, ea, sz, val):
 299         self.mem.memassign(ea, sz, val)
 300
 301     def prep_namespace(self, formname, op_fields):
 302         # TODO: get field names from form in decoder*1* (not decoder2)
 303         # decoder2 is hand-created, and decoder1.sigform is auto-generated
 304         # from spec
 305         # then "yield" fields only from op_fields rather than hard-coded
 306         # list, here.
 307         fields = self.decoder.sigforms[formname]
 308         for name in op_fields:
 309             if name == 'spr':
 310                 sig = getattr(fields, name.upper())
 311             else:
 312                 sig = getattr(fields, name)
 313             val = yield sig
 314             if name in ['BF', 'BFA']:
 315                 self.namespace[name] = val
 316             else:
 317                 self.namespace[name] = SelectableInt(val, sig.width)
 318
 319         self.namespace['XER'] = self.spr['XER']
 320         self.namespace['CA'] = self.spr['XER'][XER_bits['CA']].value
 321         self.namespace['CA32'] = self.spr['XER'][XER_bits['CA32']].value
 322
 323     def handle_carry_(self, inputs, outputs, already_done):
 324         inv_a = yield self.dec2.e.invert_a
 325         if inv_a:
 326             inputs[0] = ~inputs[0]
 327
 328         imm_ok = yield self.dec2.e.imm_data.ok
 329         if imm_ok:
 330             imm = yield self.dec2.e.imm_data.data
 331             inputs.append(SelectableInt(imm, 64))
 332         assert len(outputs) >= 1
 333         output = outputs[0]
 334         gts = [(x > output) for x in inputs]
 335         print(gts)
 336         cy = 1 if any(gts) else 0
 337         if not (1 & already_done):
 338             self.spr['XER'][XER_bits['CA']] = cy
 339
 340         print ("inputs", inputs)
 341         # 32 bit carry
 342         gts = [(x[32:64] > output[32:64]) == SelectableInt(1, 1)
 343                for x in inputs]
 344         cy32 = 1 if any(gts) else 0
 345         if not (2 & already_done):
 346             self.spr['XER'][XER_bits['CA32']] = cy32
 347
 348     def handle_overflow(self, inputs, outputs):
 349         inv_a = yield self.dec2.e.invert_a
 350         if inv_a:
 351             inputs[0] = ~inputs[0]
 352
 353         imm_ok = yield self.dec2.e.imm_data.ok
 354         if imm_ok:
 355             imm = yield self.dec2.e.imm_data.data
 356             inputs.append(SelectableInt(imm, 64))
 357         assert len(outputs) >= 1
 358         print ("handle_overflow", inputs, outputs)
 359         if len(inputs) >= 2:
 360             output = outputs[0]
 361
 362             # OV (64-bit)
 363             input_sgn = [exts(x.value, x.bits) < 0 for x in inputs]
 364             output_sgn = exts(output.value, output.bits) < 0
 365             ov = 1 if input_sgn[0] == input_sgn[1] and \
 366                 output_sgn != input_sgn[0] else 0
 367
 368             # OV (32-bit)
 369             input32_sgn = [exts(x.value, 32) < 0 for x in inputs]
 370             output32_sgn = exts(output.value, 32) < 0
 371             ov32 = 1 if input32_sgn[0] == input32_sgn[1] and \
 372                 output32_sgn != input32_sgn[0] else 0
 373
 374             self.spr['XER'][XER_bits['OV']] = ov
 375             self.spr['XER'][XER_bits['OV32']] = ov32
 376             so = self.spr['XER'][XER_bits['SO']]
 377             so = so | ov
 378             self.spr['XER'][XER_bits['SO']] = so
 379
 380     def handle_comparison(self, outputs):
 381         out = outputs[0]
 382         out = exts(out.value, out.bits)
 383         zero = SelectableInt(out == 0, 1)
 384         positive = SelectableInt(out > 0, 1)
 385         negative = SelectableInt(out < 0, 1)
 386         SO = self.spr['XER'][XER_bits['SO']]
 387         cr_field = selectconcat(negative, positive, zero, SO)
 388         self.crl[0].eq(cr_field)
 389
 390     def set_pc(self, pc_val):
 391         self.namespace['NIA'] = SelectableInt(pc_val, 64)
 392         self.pc.update(self.namespace)
 393
 394     def execute_one(self):
 395         if self.respect_pc:
 396             pc = self.pc.CIA.value
 397         else:
 398             pc = self.fake_pc
 399             self.fake_pc += 4
 400         ins = yield self.imem.ld(pc, 4, False)
 401         yield self.pdecode2.dec.raw_opcode_in.eq(ins)
 402         yield self.pdecode2.dec.bigendian.eq(0)  # little / big?
 403         opname = code.split(' ')[0]
 404         yield from call(opname)
 405
 406     def call(self, name):
 407         # TODO, asmregs is from the spec, e.g. add RT,RA,RB
 408         # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
 409         info = self.instrs[name]
 410         yield from self.prep_namespace(info.form, info.op_fields)
 411
 412         # preserve order of register names
 413         input_names = create_args(list(info.read_regs) + list(info.uninit_regs))
 414         print(input_names)
 415
 416         # main registers (RT, RA ...)
 417         inputs = []
 418         for name in input_names:
 419             regnum = yield getattr(self.decoder, name)
 420             regname = "_" + name
 421             self.namespace[regname] = regnum
 422             print('reading reg %d' % regnum)
 423             inputs.append(self.gpr(regnum))
 424
 425         # "special" registers
 426         for special in info.special_regs:
 427             if special in special_sprs:
 428                 inputs.append(self.spr[special])
 429             else:
 430                 inputs.append(self.namespace[special])
 431
 432         print(inputs)
 433         results = info.func(self, *inputs)
 434         print(results)
 435
 436         # detect if CA/CA32 already in outputs (sra*, basically)
 437         already_done = 0
 438         if info.write_regs:
 439             output_names = create_args(info.write_regs)
 440             for name in output_names:
 441                 if name == 'CA':
 442                     already_done |= 1
 443                 if name == 'CA32':
 444                     already_done |= 2
 445
 446         print ("carry already done?", bin(already_done))
 447         carry_en = yield self.dec2.e.output_carry
 448         if carry_en:
 449             yield from self.handle_carry_(inputs, results, already_done)
 450         ov_en = yield self.dec2.e.oe.oe
 451         ov_ok = yield self.dec2.e.oe.ok
 452         if ov_en & ov_ok:
 453             yield from self.handle_overflow(inputs, results)
 454         rc_en = yield self.dec2.e.rc.data
 455         if rc_en:
 456             self.handle_comparison(results)
 457
 458         # any modified return results?
 459         if info.write_regs:
 460             for name, output in zip(output_names, results):
 461                 if isinstance(output, int):
 462                     output = SelectableInt(output, 256)
 463                 if name in ['CA', 'CA32']:
 464                     if carry_en:
 465                         print ("writing %s to XER" % name, output)
 466                         self.spr['XER'][XER_bits[name]] = output.value
 467                     else:
 468                         print ("NOT writing %s to XER" % name, output)
 469                 elif name in info.special_regs:
 470                     print('writing special %s' % name, output, special_sprs)
 471                     if name in special_sprs:
 472                         self.spr[name] = output
 473                     else:
 474                         self.namespace[name].eq(output)
 475                 else:
 476                     regnum = yield getattr(self.decoder, name)
 477                     print('writing reg %d %s' % (regnum, str(output)))
 478                     if output.bits > 64:
 479                         output = SelectableInt(output.value, 64)
 480                     self.gpr[regnum] = output
 481
 482         # update program counter
 483         self.pc.update(self.namespace)
 484
 485
 486 def inject():
 487     """Decorator factory.
 488
 489     this decorator will "inject" variables into the function's namespace,
 490     from the *dictionary* in self.namespace.  it therefore becomes possible
 491     to make it look like a whole stack of variables which would otherwise
 492     need "self." inserted in front of them (*and* for those variables to be
 493     added to the instance) "appear" in the function.
 494
 495     "self.namespace['SI']" for example becomes accessible as just "SI" but
 496     *only* inside the function, when decorated.
 497     """
 498     def variable_injector(func):
 499         @wraps(func)
 500         def decorator(*args, **kwargs):
 501             try:
 502                 func_globals = func.__globals__  # Python 2.6+
 503             except AttributeError:
 504                 func_globals = func.func_globals  # Earlier versions.
 505
 506             context = args[0].namespace # variables to be injected
 507             saved_values = func_globals.copy()  # Shallow copy of dict.
 508             func_globals.update(context)
 509             result = func(*args, **kwargs)
 510             args[0].namespace = func_globals
 511             #exec (func.__code__, func_globals)
 512
 513             #finally:
 514             #    func_globals = saved_values  # Undo changes.
 515
 516             return result
 517
 518         return decorator
 519
 520     return variable_injector
 521