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, insns, InternalOp
  13 from soc.decoder.helpers import exts, trunc_div, trunc_rem
  14 from collections import namedtuple
  15 import math
  16 import sys
  17
  18 instruction_info = namedtuple('instruction_info',
  19                               'func read_regs uninit_regs write_regs ' + \
  20                               'special_regs op_fields form asmregs')
  21
  22 special_sprs = {
  23     'LR': 8,
  24     'CTR': 9,
  25     'TAR': 815,
  26     'XER': 1,
  27     'VRSAVE': 256}
  28
  29
  30 def swap_order(x, nbytes):
  31     x = x.to_bytes(nbytes, byteorder='little')
  32     x = int.from_bytes(x, byteorder='big', signed=False)
  33     return x
  34
  35
  36 def create_args(reglist, extra=None):
  37     args = OrderedSet()
  38     for reg in reglist:
  39         args.add(reg)
  40     args = list(args)
  41     if extra:
  42         args = [extra] + args
  43     return args
  44
  45
  46 class Mem:
  47
  48     def __init__(self, row_bytes=8, initial_mem=None):
  49         self.mem = {}
  50         self.bytes_per_word = row_bytes
  51         self.word_log2 = math.ceil(math.log2(row_bytes))
  52         print ("Sim-Mem", initial_mem, self.bytes_per_word, self.word_log2)
  53         if not initial_mem:
  54             return
  55
  56         # different types of memory data structures recognised (for convenience)
  57         if isinstance(initial_mem, list):
  58             initial_mem = (0, initial_mem)
  59         if isinstance(initial_mem, tuple):
  60             startaddr, mem = initial_mem
  61             initial_mem = {}
  62             for i, val in enumerate(mem):
  63                 initial_mem[startaddr + row_bytes*i] = (val, row_bytes)
  64
  65         for addr, (val, width) in initial_mem.items():
  66             #val = swap_order(val, width)
  67             self.st(addr, val, width, swap=False)
  68
  69     def _get_shifter_mask(self, wid, remainder):
  70         shifter = ((self.bytes_per_word - wid) - remainder) * \
  71             8  # bits per byte
  72         # XXX https://bugs.libre-soc.org/show_bug.cgi?id=377
  73         # BE/LE mode?
  74         shifter = remainder * 8
  75         mask = (1 << (wid * 8)) - 1
  76         print ("width,rem,shift,mask", wid, remainder, hex(shifter), hex(mask))
  77         return shifter, mask
  78
  79     # TODO: Implement ld/st of lesser width
  80     def ld(self, address, width=8, swap=True, check_in_mem=False):
  81         print("ld from addr 0x{:x} width {:d}".format(address, width))
  82         remainder = address & (self.bytes_per_word - 1)
  83         address = address >> self.word_log2
  84         assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
  85         if address in self.mem:
  86             val = self.mem[address]
  87         elif check_in_mem:
  88             return None
  89         else:
  90             val = 0
  91         print("mem @ 0x{:x} rem {:d} : 0x{:x}".format(address, remainder, val))
  92
  93         if width != self.bytes_per_word:
  94             shifter, mask = self._get_shifter_mask(width, remainder)
  95             print ("masking", hex(val), hex(mask<<shifter), shifter)
  96             val = val & (mask << shifter)
  97             val >>= shifter
  98         if swap:
  99             val = swap_order(val, width)
 100         print("Read 0x{:x} from addr 0x{:x}".format(val, address))
 101         return val
 102
 103     def st(self, addr, v, width=8, swap=True):
 104         staddr = addr
 105         remainder = addr & (self.bytes_per_word - 1)
 106         addr = addr >> self.word_log2
 107         print("Writing 0x{:x} to ST 0x{:x} memaddr 0x{:x}/{:x}".format(v,
 108                         staddr, addr, remainder, swap))
 109         assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
 110         if swap:
 111             v = swap_order(v, width)
 112         if width != self.bytes_per_word:
 113             if addr in self.mem:
 114                 val = self.mem[addr]
 115             else:
 116                 val = 0
 117             shifter, mask = self._get_shifter_mask(width, remainder)
 118             val &= ~(mask << shifter)
 119             val |= v << shifter
 120             self.mem[addr] = val
 121         else:
 122             self.mem[addr] = v
 123         print("mem @ 0x{:x}: 0x{:x}".format(addr, self.mem[addr]))
 124
 125     def __call__(self, addr, sz):
 126         val = self.ld(addr.value, sz)
 127         print ("memread", addr, sz, val)
 128         return SelectableInt(val, sz*8)
 129
 130     def memassign(self, addr, sz, val):
 131         print ("memassign", addr, sz, val)
 132         self.st(addr.value, val.value, sz)
 133
 134
 135 class GPR(dict):
 136     def __init__(self, decoder, regfile):
 137         dict.__init__(self)
 138         self.sd = decoder
 139         for i in range(32):
 140             self[i] = SelectableInt(regfile[i], 64)
 141
 142     def __call__(self, ridx):
 143         return self[ridx]
 144
 145     def set_form(self, form):
 146         self.form = form
 147
 148     def getz(self, rnum):
 149         #rnum = rnum.value # only SelectableInt allowed
 150         print("GPR getzero", rnum)
 151         if rnum == 0:
 152             return SelectableInt(0, 64)
 153         return self[rnum]
 154
 155     def _get_regnum(self, attr):
 156         getform = self.sd.sigforms[self.form]
 157         rnum = getattr(getform, attr)
 158         return rnum
 159
 160     def ___getitem__(self, attr):
 161         print("GPR getitem", attr)
 162         rnum = self._get_regnum(attr)
 163         return self.regfile[rnum]
 164
 165     def dump(self):
 166         for i in range(0, len(self), 8):
 167             s = []
 168             for j in range(8):
 169                 s.append("%08x" % self[i+j].value)
 170             s = ' '.join(s)
 171             print("reg", "%2d" % i, s)
 172
 173 class PC:
 174     def __init__(self, pc_init=0):
 175         self.CIA = SelectableInt(pc_init, 64)
 176         self.NIA = self.CIA + SelectableInt(4, 64)
 177
 178     def update(self, namespace):
 179         self.CIA = namespace['NIA'].narrow(64)
 180         self.NIA = self.CIA + SelectableInt(4, 64)
 181         namespace['CIA'] = self.CIA
 182         namespace['NIA'] = self.NIA
 183
 184
 185 class SPR(dict):
 186     def __init__(self, dec2, initial_sprs={}):
 187         self.sd = dec2
 188         dict.__init__(self)
 189         self.update(initial_sprs)
 190
 191     def __getitem__(self, key):
 192         # if key in special_sprs get the special spr, otherwise return key
 193         if isinstance(key, SelectableInt):
 194             key = key.value
 195         key = special_sprs.get(key, key)
 196         if key in self:
 197             return dict.__getitem__(self, key)
 198         else:
 199             info = spr_dict[key]
 200             dict.__setitem__(self, key, SelectableInt(0, info.length))
 201             return dict.__getitem__(self, key)
 202
 203     def __setitem__(self, key, value):
 204         if isinstance(key, SelectableInt):
 205             key = key.value
 206         key = special_sprs.get(key, key)
 207         dict.__setitem__(self, key, value)
 208
 209     def __call__(self, ridx):
 210         return self[ridx]
 211
 212
 213 class ISACaller:
 214     # decoder2 - an instance of power_decoder2
 215     # regfile - a list of initial values for the registers
 216     # initial_{etc} - initial values for SPRs, Condition Register, Mem, MSR
 217     # respect_pc - tracks the program counter.  requires initial_insns
 218     def __init__(self, decoder2, regfile, initial_sprs=None, initial_cr=0,
 219                        initial_mem=None, initial_msr=0,
 220                        initial_insns=None, respect_pc=False,
 221                        disassembly=None):
 222
 223         self.respect_pc = respect_pc
 224         if initial_sprs is None:
 225             initial_sprs = {}
 226         if initial_mem is None:
 227             initial_mem = {}
 228         if initial_insns is None:
 229             initial_insns = {}
 230             assert self.respect_pc == False, "instructions required to honor pc"
 231
 232         print ("ISACaller insns", respect_pc, initial_insns, disassembly)
 233
 234         # "fake program counter" mode (for unit testing)
 235         self.fake_pc = 0
 236         if not respect_pc:
 237             if isinstance(initial_mem, tuple):
 238                 self.fake_pc = initial_mem[0]
 239
 240         # disassembly: we need this for now (not given from the decoder)
 241         self.disassembly = {}
 242         if disassembly:
 243             for i, code in enumerate(disassembly):
 244                 self.disassembly[i*4 + self.fake_pc] = code
 245
 246         # set up registers, instruction memory, data memory, PC, SPRs, MSR
 247         self.gpr = GPR(decoder2, regfile)
 248         self.mem = Mem(row_bytes=8, initial_mem=initial_mem)
 249         self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
 250         self.pc = PC()
 251         self.spr = SPR(decoder2, initial_sprs)
 252         self.msr = SelectableInt(initial_msr, 64) # underlying reg
 253
 254         # TODO, needed here:
 255         # FPR (same as GPR except for FP nums)
 256         # 4.2.2 p124 FPSCR (definitely "separate" - not in SPR)
 257         #            note that mffs, mcrfs, mtfsf "manage" this FPSCR
 258         # 2.3.1 CR (and sub-fields CR0..CR6 - CR0 SO comes from XER.SO)
 259         #         note that mfocrf, mfcr, mtcr, mtocrf, mcrxrx "manage" CRs
 260         #         -- Done
 261         # 2.3.2 LR   (actually SPR #8) -- Done
 262         # 2.3.3 CTR  (actually SPR #9) -- Done
 263         # 2.3.4 TAR  (actually SPR #815)
 264         # 3.2.2 p45 XER  (actually SPR #1) -- Done
 265         # 3.2.3 p46 p232 VRSAVE (actually SPR #256)
 266
 267         # create CR then allow portions of it to be "selectable" (below)
 268         self._cr = SelectableInt(initial_cr, 64) # underlying reg
 269         self.cr = FieldSelectableInt(self._cr, list(range(32,64)))
 270
 271         # "undefined", just set to variable-bit-width int (use exts "max")
 272         self.undefined = SelectableInt(0, 256) # TODO, not hard-code 256!
 273
 274         self.namespace = {'GPR': self.gpr,
 275                           'MEM': self.mem,
 276                           'SPR': self.spr,
 277                           'memassign': self.memassign,
 278                           'NIA': self.pc.NIA,
 279                           'CIA': self.pc.CIA,
 280                           'CR': self.cr,
 281                           'MSR': self.msr,
 282                           'undefined': self.undefined,
 283                           'mode_is_64bit': True,
 284                           'SO': XER_bits['SO']
 285                           }
 286
 287         # field-selectable versions of Condition Register TODO check bitranges?
 288         self.crl = []
 289         for i in range(8):
 290             bits = tuple(range(i*4, (i+1)*4))# errr... maybe?
 291             _cr = FieldSelectableInt(self.cr, bits)
 292             self.crl.append(_cr)
 293             self.namespace["CR%d" % i] = _cr
 294
 295         self.decoder = decoder2.dec
 296         self.dec2 = decoder2
 297
 298     def TRAP(self, trap_addr=0x700):
 299         print ("TRAP: TODO")
 300         #self.namespace['NIA'] = trap_addr
 301         #self.SRR0 = self.namespace['CIA'] + 4
 302         #self.SRR1 = self.namespace['MSR']
 303         #self.namespace['MSR'][45] = 1
 304         # store CIA(+4?) in SRR0, set NIA to 0x700
 305         # store MSR in SRR1, set MSR to um errr something, have to check spec
 306
 307     def memassign(self, ea, sz, val):
 308         self.mem.memassign(ea, sz, val)
 309
 310     def prep_namespace(self, formname, op_fields):
 311         # TODO: get field names from form in decoder*1* (not decoder2)
 312         # decoder2 is hand-created, and decoder1.sigform is auto-generated
 313         # from spec
 314         # then "yield" fields only from op_fields rather than hard-coded
 315         # list, here.
 316         fields = self.decoder.sigforms[formname]
 317         for name in op_fields:
 318             if name == 'spr':
 319                 sig = getattr(fields, name.upper())
 320             else:
 321                 sig = getattr(fields, name)
 322             val = yield sig
 323             if name in ['BF', 'BFA']:
 324                 self.namespace[name] = val
 325             else:
 326                 self.namespace[name] = SelectableInt(val, sig.width)
 327
 328         self.namespace['XER'] = self.spr['XER']
 329         self.namespace['CA'] = self.spr['XER'][XER_bits['CA']].value
 330         self.namespace['CA32'] = self.spr['XER'][XER_bits['CA32']].value
 331
 332     def handle_carry_(self, inputs, outputs, already_done):
 333         inv_a = yield self.dec2.e.invert_a
 334         if inv_a:
 335             inputs[0] = ~inputs[0]
 336
 337         imm_ok = yield self.dec2.e.imm_data.ok
 338         if imm_ok:
 339             imm = yield self.dec2.e.imm_data.data
 340             inputs.append(SelectableInt(imm, 64))
 341         assert len(outputs) >= 1
 342         print ("outputs", repr(outputs))
 343         if isinstance(outputs, list) or isinstance(outputs, tuple):
 344             output = outputs[0]
 345         else:
 346             output = outputs
 347         gts = []
 348         for x in inputs:
 349             print ("gt input", x, output)
 350             gt = (x > output)
 351             gts.append(gt)
 352         print(gts)
 353         cy = 1 if any(gts) else 0
 354         if not (1 & already_done):
 355             self.spr['XER'][XER_bits['CA']] = cy
 356
 357         print ("inputs", inputs)
 358         # 32 bit carry
 359         gts = []
 360         for x in inputs:
 361             print ("input", x, output)
 362             gt = (x[32:64] > output[32:64]) == SelectableInt(1, 1)
 363             gts.append(gt)
 364         cy32 = 1 if any(gts) else 0
 365         if not (2 & already_done):
 366             self.spr['XER'][XER_bits['CA32']] = cy32
 367
 368     def handle_overflow(self, inputs, outputs, div_overflow):
 369         inv_a = yield self.dec2.e.invert_a
 370         if inv_a:
 371             inputs[0] = ~inputs[0]
 372
 373         imm_ok = yield self.dec2.e.imm_data.ok
 374         if imm_ok:
 375             imm = yield self.dec2.e.imm_data.data
 376             inputs.append(SelectableInt(imm, 64))
 377         assert len(outputs) >= 1
 378         print ("handle_overflow", inputs, outputs, div_overflow)
 379         if len(inputs) < 2 and div_overflow != 1:
 380             return
 381
 382         # div overflow is different: it's returned by the pseudo-code
 383         # because it's more complex than can be done by analysing the output
 384         if div_overflow == 1:
 385             ov, ov32 = 1, 1
 386         # arithmetic overflow can be done by analysing the input and output
 387         elif len(inputs) >= 2:
 388             output = outputs[0]
 389
 390             # OV (64-bit)
 391             input_sgn = [exts(x.value, x.bits) < 0 for x in inputs]
 392             output_sgn = exts(output.value, output.bits) < 0
 393             ov = 1 if input_sgn[0] == input_sgn[1] and \
 394                 output_sgn != input_sgn[0] else 0
 395
 396             # OV (32-bit)
 397             input32_sgn = [exts(x.value, 32) < 0 for x in inputs]
 398             output32_sgn = exts(output.value, 32) < 0
 399             ov32 = 1 if input32_sgn[0] == input32_sgn[1] and \
 400                 output32_sgn != input32_sgn[0] else 0
 401
 402         self.spr['XER'][XER_bits['OV']] = ov
 403         self.spr['XER'][XER_bits['OV32']] = ov32
 404         so = self.spr['XER'][XER_bits['SO']]
 405         so = so | ov
 406         self.spr['XER'][XER_bits['SO']] = so
 407
 408     def handle_comparison(self, outputs):
 409         out = outputs[0]
 410         out = exts(out.value, out.bits)
 411         zero = SelectableInt(out == 0, 1)
 412         positive = SelectableInt(out > 0, 1)
 413         negative = SelectableInt(out < 0, 1)
 414         SO = self.spr['XER'][XER_bits['SO']]
 415         cr_field = selectconcat(negative, positive, zero, SO)
 416         self.crl[0].eq(cr_field)
 417
 418     def set_pc(self, pc_val):
 419         self.namespace['NIA'] = SelectableInt(pc_val, 64)
 420         self.pc.update(self.namespace)
 421
 422     def setup_one(self):
 423         """set up one instruction
 424         """
 425         if self.respect_pc:
 426             pc = self.pc.CIA.value
 427         else:
 428             pc = self.fake_pc
 429         self._pc = pc
 430         ins = self.imem.ld(pc, 4, False, True)
 431         if ins is None:
 432             raise KeyError("no instruction at 0x%x" % pc)
 433         print("setup: 0x%x 0x%x %s" % (pc, ins & 0xffffffff, bin(ins)))
 434         print ("NIA, CIA", self.pc.CIA.value, self.pc.NIA.value)
 435
 436         yield self.dec2.dec.raw_opcode_in.eq(ins & 0xffffffff)
 437         yield self.dec2.dec.bigendian.eq(0)  # little / big?
 438
 439     def execute_one(self):
 440         """execute one instruction
 441         """
 442         # get the disassembly code for this instruction
 443         code = self.disassembly[self._pc]
 444         print("sim-execute", hex(self._pc), code)
 445         opname = code.split(' ')[0]
 446         yield from self.call(opname)
 447
 448         if not self.respect_pc:
 449             self.fake_pc += 4
 450         print ("NIA, CIA", self.pc.CIA.value, self.pc.NIA.value)
 451
 452     def get_assembly_name(self):
 453         # TODO, asmregs is from the spec, e.g. add RT,RA,RB
 454         # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
 455         asmcode = yield self.dec2.dec.op.asmcode
 456         asmop = insns.get(asmcode, None)
 457
 458         # sigh reconstruct the assembly instruction name
 459         ov_en = yield self.dec2.e.oe.oe
 460         ov_ok = yield self.dec2.e.oe.ok
 461         if ov_en & ov_ok:
 462             asmop += "."
 463         lk = yield self.dec2.e.lk
 464         if lk:
 465             asmop += "l"
 466         int_op = yield self.dec2.dec.op.internal_op
 467         print ("int_op", int_op)
 468         if int_op in [InternalOp.OP_B.value, InternalOp.OP_BC.value]:
 469             AA = yield self.dec2.dec.fields.FormI.AA[0:-1]
 470             print ("AA", AA)
 471             if AA:
 472                 asmop += "a"
 473         if int_op == InternalOp.OP_MFCR.value:
 474             dec_insn = yield self.dec2.e.insn
 475             if dec_insn & (1<<20) != 0: # sigh
 476                 asmop = 'mfocrf'
 477             else:
 478                 asmop = 'mfcr'
 479         # XXX TODO: for whatever weird reason this doesn't work
 480         # https://bugs.libre-soc.org/show_bug.cgi?id=390
 481         if int_op == InternalOp.OP_MTCRF.value:
 482             dec_insn = yield self.dec2.e.insn
 483             if dec_insn & (1<<20) != 0: # sigh
 484                 asmop = 'mtocrf'
 485             else:
 486                 asmop = 'mtcrf'
 487         return asmop
 488
 489     def call(self, name):
 490         # TODO, asmregs is from the spec, e.g. add RT,RA,RB
 491         # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
 492         asmop = yield from self.get_assembly_name()
 493         print  ("call", name, asmop)
 494         if name not in ['mtcrf', 'mtocrf']:
 495             assert name == asmop, "name %s != %s" % (name, asmop)
 496
 497         info = self.instrs[name]
 498         yield from self.prep_namespace(info.form, info.op_fields)
 499
 500         # preserve order of register names
 501         input_names = create_args(list(info.read_regs) + list(info.uninit_regs))
 502         print(input_names)
 503
 504         # main registers (RT, RA ...)
 505         inputs = []
 506         for name in input_names:
 507             regnum = yield getattr(self.decoder, name)
 508             regname = "_" + name
 509             self.namespace[regname] = regnum
 510             print('reading reg %d' % regnum)
 511             inputs.append(self.gpr(regnum))
 512
 513         # "special" registers
 514         for special in info.special_regs:
 515             if special in special_sprs:
 516                 inputs.append(self.spr[special])
 517             else:
 518                 inputs.append(self.namespace[special])
 519
 520         print(inputs)
 521         results = info.func(self, *inputs)
 522         print(results)
 523
 524         # detect if CA/CA32 already in outputs (sra*, basically)
 525         already_done = 0
 526         if info.write_regs:
 527             output_names = create_args(info.write_regs)
 528             for name in output_names:
 529                 if name == 'CA':
 530                     already_done |= 1
 531                 if name == 'CA32':
 532                     already_done |= 2
 533
 534         print ("carry already done?", bin(already_done))
 535         carry_en = yield self.dec2.e.output_carry
 536         if carry_en:
 537             yield from self.handle_carry_(inputs, results, already_done)
 538
 539         # detect if overflow was in return result
 540         overflow = None
 541         if info.write_regs:
 542             for name, output in zip(output_names, results):
 543                 if name == 'overflow':
 544                     overflow = output
 545
 546         ov_en = yield self.dec2.e.oe.oe
 547         ov_ok = yield self.dec2.e.oe.ok
 548         print ("internal overflow", overflow)
 549         if ov_en & ov_ok:
 550             yield from self.handle_overflow(inputs, results, overflow)
 551
 552         rc_en = yield self.dec2.e.rc.data
 553         if rc_en:
 554             self.handle_comparison(results)
 555
 556         # any modified return results?
 557         if info.write_regs:
 558             for name, output in zip(output_names, results):
 559                 if name == 'overflow': # ignore, done already (above)
 560                     continue
 561                 if isinstance(output, int):
 562                     output = SelectableInt(output, 256)
 563                 if name in ['CA', 'CA32']:
 564                     if carry_en:
 565                         print ("writing %s to XER" % name, output)
 566                         self.spr['XER'][XER_bits[name]] = output.value
 567                     else:
 568                         print ("NOT writing %s to XER" % name, output)
 569                 elif name in info.special_regs:
 570                     print('writing special %s' % name, output, special_sprs)
 571                     if name in special_sprs:
 572                         self.spr[name] = output
 573                     else:
 574                         self.namespace[name].eq(output)
 575                 else:
 576                     regnum = yield getattr(self.decoder, name)
 577                     print('writing reg %d %s' % (regnum, str(output)))
 578                     if output.bits > 64:
 579                         output = SelectableInt(output.value, 64)
 580                     self.gpr[regnum] = output
 581
 582         # update program counter
 583         self.pc.update(self.namespace)
 584
 585
 586 def inject():
 587     """Decorator factory.
 588
 589     this decorator will "inject" variables into the function's namespace,
 590     from the *dictionary* in self.namespace.  it therefore becomes possible
 591     to make it look like a whole stack of variables which would otherwise
 592     need "self." inserted in front of them (*and* for those variables to be
 593     added to the instance) "appear" in the function.
 594
 595     "self.namespace['SI']" for example becomes accessible as just "SI" but
 596     *only* inside the function, when decorated.
 597     """
 598     def variable_injector(func):
 599         @wraps(func)
 600         def decorator(*args, **kwargs):
 601             try:
 602                 func_globals = func.__globals__  # Python 2.6+
 603             except AttributeError:
 604                 func_globals = func.func_globals  # Earlier versions.
 605
 606             context = args[0].namespace # variables to be injected
 607             saved_values = func_globals.copy()  # Shallow copy of dict.
 608             func_globals.update(context)
 609             result = func(*args, **kwargs)
 610             args[0].namespace = func_globals
 611             #exec (func.__code__, func_globals)
 612
 613             #finally:
 614             #    func_globals = saved_values  # Undo changes.
 615
 616             return result
 617
 618         return decorator
 619
 620     return variable_injector
 621