273ffd17b9814047517559a0174c6c1c8bf9cb23
[soc.git] / 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, spr_byname, XER_bits,
13 insns, InternalOp)
14 from soc.decoder.helpers import exts, trunc_div, trunc_rem
15 from soc.consts import PI, MSR
16
17 from collections import namedtuple
18 import math
19 import sys
20
21 instruction_info = namedtuple('instruction_info',
22 'func read_regs uninit_regs write_regs ' + \
23 'special_regs op_fields form asmregs')
24
25 special_sprs = {
26 'LR': 8,
27 'CTR': 9,
28 'TAR': 815,
29 'XER': 1,
30 'VRSAVE': 256}
31
32
33 def swap_order(x, nbytes):
34 x = x.to_bytes(nbytes, byteorder='little')
35 x = int.from_bytes(x, byteorder='big', signed=False)
36 return x
37
38
39 def create_args(reglist, extra=None):
40 args = OrderedSet()
41 for reg in reglist:
42 args.add(reg)
43 args = list(args)
44 if extra:
45 args = [extra] + args
46 return args
47
48
49 class Mem:
50
51 def __init__(self, row_bytes=8, initial_mem=None):
52 self.mem = {}
53 self.bytes_per_word = row_bytes
54 self.word_log2 = math.ceil(math.log2(row_bytes))
55 print ("Sim-Mem", initial_mem, self.bytes_per_word, self.word_log2)
56 if not initial_mem:
57 return
58
59 # different types of memory data structures recognised (for convenience)
60 if isinstance(initial_mem, list):
61 initial_mem = (0, initial_mem)
62 if isinstance(initial_mem, tuple):
63 startaddr, mem = initial_mem
64 initial_mem = {}
65 for i, val in enumerate(mem):
66 initial_mem[startaddr + row_bytes*i] = (val, row_bytes)
67
68 for addr, (val, width) in initial_mem.items():
69 #val = swap_order(val, width)
70 self.st(addr, val, width, swap=False)
71
72 def _get_shifter_mask(self, wid, remainder):
73 shifter = ((self.bytes_per_word - wid) - remainder) * \
74 8 # bits per byte
75 # XXX https://bugs.libre-soc.org/show_bug.cgi?id=377
76 # BE/LE mode?
77 shifter = remainder * 8
78 mask = (1 << (wid * 8)) - 1
79 print ("width,rem,shift,mask", wid, remainder, hex(shifter), hex(mask))
80 return shifter, mask
81
82 # TODO: Implement ld/st of lesser width
83 def ld(self, address, width=8, swap=True, check_in_mem=False):
84 print("ld from addr 0x{:x} width {:d}".format(address, width))
85 remainder = address & (self.bytes_per_word - 1)
86 address = address >> self.word_log2
87 assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
88 if address in self.mem:
89 val = self.mem[address]
90 elif check_in_mem:
91 return None
92 else:
93 val = 0
94 print("mem @ 0x{:x} rem {:d} : 0x{:x}".format(address, remainder, val))
95
96 if width != self.bytes_per_word:
97 shifter, mask = self._get_shifter_mask(width, remainder)
98 print ("masking", hex(val), hex(mask<<shifter), shifter)
99 val = val & (mask << shifter)
100 val >>= shifter
101 if swap:
102 val = swap_order(val, width)
103 print("Read 0x{:x} from addr 0x{:x}".format(val, address))
104 return val
105
106 def st(self, addr, v, width=8, swap=True):
107 staddr = addr
108 remainder = addr & (self.bytes_per_word - 1)
109 addr = addr >> self.word_log2
110 print("Writing 0x{:x} to ST 0x{:x} memaddr 0x{:x}/{:x}".format(v,
111 staddr, addr, remainder, swap))
112 assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
113 if swap:
114 v = swap_order(v, width)
115 if width != self.bytes_per_word:
116 if addr in self.mem:
117 val = self.mem[addr]
118 else:
119 val = 0
120 shifter, mask = self._get_shifter_mask(width, remainder)
121 val &= ~(mask << shifter)
122 val |= v << shifter
123 self.mem[addr] = val
124 else:
125 self.mem[addr] = v
126 print("mem @ 0x{:x}: 0x{:x}".format(addr, self.mem[addr]))
127
128 def __call__(self, addr, sz):
129 val = self.ld(addr.value, sz)
130 print ("memread", addr, sz, val)
131 return SelectableInt(val, sz*8)
132
133 def memassign(self, addr, sz, val):
134 print ("memassign", addr, sz, val)
135 self.st(addr.value, val.value, sz)
136
137
138 class GPR(dict):
139 def __init__(self, decoder, regfile):
140 dict.__init__(self)
141 self.sd = decoder
142 for i in range(32):
143 self[i] = SelectableInt(regfile[i], 64)
144
145 def __call__(self, ridx):
146 return self[ridx]
147
148 def set_form(self, form):
149 self.form = form
150
151 def getz(self, rnum):
152 #rnum = rnum.value # only SelectableInt allowed
153 print("GPR getzero", rnum)
154 if rnum == 0:
155 return SelectableInt(0, 64)
156 return self[rnum]
157
158 def _get_regnum(self, attr):
159 getform = self.sd.sigforms[self.form]
160 rnum = getattr(getform, attr)
161 return rnum
162
163 def ___getitem__(self, attr):
164 print("GPR getitem", attr)
165 rnum = self._get_regnum(attr)
166 return self.regfile[rnum]
167
168 def dump(self):
169 for i in range(0, len(self), 8):
170 s = []
171 for j in range(8):
172 s.append("%08x" % self[i+j].value)
173 s = ' '.join(s)
174 print("reg", "%2d" % i, s)
175
176 class PC:
177 def __init__(self, pc_init=0):
178 self.CIA = SelectableInt(pc_init, 64)
179 self.NIA = self.CIA + SelectableInt(4, 64)
180
181 def update(self, namespace):
182 self.CIA = namespace['NIA'].narrow(64)
183 self.NIA = self.CIA + SelectableInt(4, 64)
184 namespace['CIA'] = self.CIA
185 namespace['NIA'] = self.NIA
186
187
188 class SPR(dict):
189 def __init__(self, dec2, initial_sprs={}):
190 self.sd = dec2
191 dict.__init__(self)
192 for key, v in initial_sprs.items():
193 if isinstance(key, SelectableInt):
194 key = key.value
195 key = special_sprs.get(key, key)
196 if isinstance(key, int):
197 info = spr_dict[key]
198 else:
199 info = spr_byname[key]
200 if not isinstance(v, SelectableInt):
201 v = SelectableInt(v, info.length)
202 self[key] = v
203
204 def __getitem__(self, key):
205 print ("get spr", key)
206 print ("dict", self.items())
207 # if key in special_sprs get the special spr, otherwise return key
208 if isinstance(key, SelectableInt):
209 key = key.value
210 if isinstance(key, int):
211 key = spr_dict[key].SPR
212 key = special_sprs.get(key, key)
213 if key in self:
214 res = dict.__getitem__(self, key)
215 else:
216 if isinstance(key, int):
217 info = spr_dict[key]
218 else:
219 info = spr_byname[key]
220 dict.__setitem__(self, key, SelectableInt(0, info.length))
221 res = dict.__getitem__(self, key)
222 print ("spr returning", key, res)
223 return res
224
225 def __setitem__(self, key, value):
226 if isinstance(key, SelectableInt):
227 key = key.value
228 if isinstance(key, int):
229 key = spr_dict[key].SPR
230 print ("spr key", key)
231 key = special_sprs.get(key, key)
232 print ("setting spr", key, value)
233 dict.__setitem__(self, key, value)
234
235 def __call__(self, ridx):
236 return self[ridx]
237
238
239 class ISACaller:
240 # decoder2 - an instance of power_decoder2
241 # regfile - a list of initial values for the registers
242 # initial_{etc} - initial values for SPRs, Condition Register, Mem, MSR
243 # respect_pc - tracks the program counter. requires initial_insns
244 def __init__(self, decoder2, regfile, initial_sprs=None, initial_cr=0,
245 initial_mem=None, initial_msr=0,
246 initial_insns=None, respect_pc=False,
247 disassembly=None):
248
249 self.respect_pc = respect_pc
250 if initial_sprs is None:
251 initial_sprs = {}
252 if initial_mem is None:
253 initial_mem = {}
254 if initial_insns is None:
255 initial_insns = {}
256 assert self.respect_pc == False, "instructions required to honor pc"
257
258 print ("ISACaller insns", respect_pc, initial_insns, disassembly)
259
260 # "fake program counter" mode (for unit testing)
261 self.fake_pc = 0
262 if not respect_pc:
263 if isinstance(initial_mem, tuple):
264 self.fake_pc = initial_mem[0]
265
266 # disassembly: we need this for now (not given from the decoder)
267 self.disassembly = {}
268 if disassembly:
269 for i, code in enumerate(disassembly):
270 self.disassembly[i*4 + self.fake_pc] = code
271
272 # set up registers, instruction memory, data memory, PC, SPRs, MSR
273 self.gpr = GPR(decoder2, regfile)
274 self.mem = Mem(row_bytes=8, initial_mem=initial_mem)
275 self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
276 self.pc = PC()
277 self.spr = SPR(decoder2, initial_sprs)
278 self.msr = SelectableInt(initial_msr, 64) # underlying reg
279
280 # TODO, needed here:
281 # FPR (same as GPR except for FP nums)
282 # 4.2.2 p124 FPSCR (definitely "separate" - not in SPR)
283 # note that mffs, mcrfs, mtfsf "manage" this FPSCR
284 # 2.3.1 CR (and sub-fields CR0..CR6 - CR0 SO comes from XER.SO)
285 # note that mfocrf, mfcr, mtcr, mtocrf, mcrxrx "manage" CRs
286 # -- Done
287 # 2.3.2 LR (actually SPR #8) -- Done
288 # 2.3.3 CTR (actually SPR #9) -- Done
289 # 2.3.4 TAR (actually SPR #815)
290 # 3.2.2 p45 XER (actually SPR #1) -- Done
291 # 3.2.3 p46 p232 VRSAVE (actually SPR #256)
292
293 # create CR then allow portions of it to be "selectable" (below)
294 self._cr = SelectableInt(initial_cr, 64) # underlying reg
295 self.cr = FieldSelectableInt(self._cr, list(range(32,64)))
296
297 # "undefined", just set to variable-bit-width int (use exts "max")
298 self.undefined = SelectableInt(0, 256) # TODO, not hard-code 256!
299
300 self.namespace = {}
301 self.namespace.update(self.spr)
302 self.namespace.update({'GPR': self.gpr,
303 'MEM': self.mem,
304 'SPR': self.spr,
305 'memassign': self.memassign,
306 'NIA': self.pc.NIA,
307 'CIA': self.pc.CIA,
308 'CR': self.cr,
309 'MSR': self.msr,
310 'undefined': self.undefined,
311 'mode_is_64bit': True,
312 'SO': XER_bits['SO']
313 })
314
315
316 # field-selectable versions of Condition Register TODO check bitranges?
317 self.crl = []
318 for i in range(8):
319 bits = tuple(range(i*4, (i+1)*4))# errr... maybe?
320 _cr = FieldSelectableInt(self.cr, bits)
321 self.crl.append(_cr)
322 self.namespace["CR%d" % i] = _cr
323
324 self.decoder = decoder2.dec
325 self.dec2 = decoder2
326
327 def TRAP(self, trap_addr=0x700):
328 print ("TRAP: TODO")
329 # store CIA(+4?) in SRR0, set NIA to 0x700
330 # store MSR in SRR1, set MSR to um errr something, have to check spec
331 self.spr['SRR0'] = self.pc.CIA
332 self.spr['SRR1'] = self.namespace['MSR']
333 self.set_pc(trap_addr)
334 self.namespace['MSR'][63-PI.TRAP] = 1 # bit 45, "this is a trap"
335
336 def memassign(self, ea, sz, val):
337 self.mem.memassign(ea, sz, val)
338
339 def prep_namespace(self, formname, op_fields):
340 # TODO: get field names from form in decoder*1* (not decoder2)
341 # decoder2 is hand-created, and decoder1.sigform is auto-generated
342 # from spec
343 # then "yield" fields only from op_fields rather than hard-coded
344 # list, here.
345 fields = self.decoder.sigforms[formname]
346 for name in op_fields:
347 if name == 'spr':
348 sig = getattr(fields, name.upper())
349 else:
350 sig = getattr(fields, name)
351 val = yield sig
352 if name in ['BF', 'BFA']:
353 self.namespace[name] = val
354 else:
355 self.namespace[name] = SelectableInt(val, sig.width)
356
357 self.namespace['XER'] = self.spr['XER']
358 self.namespace['CA'] = self.spr['XER'][XER_bits['CA']].value
359 self.namespace['CA32'] = self.spr['XER'][XER_bits['CA32']].value
360
361 def handle_carry_(self, inputs, outputs, already_done):
362 inv_a = yield self.dec2.e.invert_a
363 if inv_a:
364 inputs[0] = ~inputs[0]
365
366 imm_ok = yield self.dec2.e.imm_data.ok
367 if imm_ok:
368 imm = yield self.dec2.e.imm_data.data
369 inputs.append(SelectableInt(imm, 64))
370 assert len(outputs) >= 1
371 print ("outputs", repr(outputs))
372 if isinstance(outputs, list) or isinstance(outputs, tuple):
373 output = outputs[0]
374 else:
375 output = outputs
376 gts = []
377 for x in inputs:
378 print ("gt input", x, output)
379 gt = (x > output)
380 gts.append(gt)
381 print(gts)
382 cy = 1 if any(gts) else 0
383 if not (1 & already_done):
384 self.spr['XER'][XER_bits['CA']] = cy
385
386 print ("inputs", inputs)
387 # 32 bit carry
388 gts = []
389 for x in inputs:
390 print ("input", x, output)
391 gt = (x[32:64] > output[32:64]) == SelectableInt(1, 1)
392 gts.append(gt)
393 cy32 = 1 if any(gts) else 0
394 if not (2 & already_done):
395 self.spr['XER'][XER_bits['CA32']] = cy32
396
397 def handle_overflow(self, inputs, outputs, div_overflow):
398 inv_a = yield self.dec2.e.invert_a
399 if inv_a:
400 inputs[0] = ~inputs[0]
401
402 imm_ok = yield self.dec2.e.imm_data.ok
403 if imm_ok:
404 imm = yield self.dec2.e.imm_data.data
405 inputs.append(SelectableInt(imm, 64))
406 assert len(outputs) >= 1
407 print ("handle_overflow", inputs, outputs, div_overflow)
408 if len(inputs) < 2 and div_overflow != 1:
409 return
410
411 # div overflow is different: it's returned by the pseudo-code
412 # because it's more complex than can be done by analysing the output
413 if div_overflow == 1:
414 ov, ov32 = 1, 1
415 # arithmetic overflow can be done by analysing the input and output
416 elif len(inputs) >= 2:
417 output = outputs[0]
418
419 # OV (64-bit)
420 input_sgn = [exts(x.value, x.bits) < 0 for x in inputs]
421 output_sgn = exts(output.value, output.bits) < 0
422 ov = 1 if input_sgn[0] == input_sgn[1] and \
423 output_sgn != input_sgn[0] else 0
424
425 # OV (32-bit)
426 input32_sgn = [exts(x.value, 32) < 0 for x in inputs]
427 output32_sgn = exts(output.value, 32) < 0
428 ov32 = 1 if input32_sgn[0] == input32_sgn[1] and \
429 output32_sgn != input32_sgn[0] else 0
430
431 self.spr['XER'][XER_bits['OV']] = ov
432 self.spr['XER'][XER_bits['OV32']] = ov32
433 so = self.spr['XER'][XER_bits['SO']]
434 so = so | ov
435 self.spr['XER'][XER_bits['SO']] = so
436
437 def handle_comparison(self, outputs):
438 out = outputs[0]
439 out = exts(out.value, out.bits)
440 zero = SelectableInt(out == 0, 1)
441 positive = SelectableInt(out > 0, 1)
442 negative = SelectableInt(out < 0, 1)
443 SO = self.spr['XER'][XER_bits['SO']]
444 cr_field = selectconcat(negative, positive, zero, SO)
445 self.crl[0].eq(cr_field)
446
447 def set_pc(self, pc_val):
448 self.namespace['NIA'] = SelectableInt(pc_val, 64)
449 self.pc.update(self.namespace)
450
451 def setup_one(self):
452 """set up one instruction
453 """
454 if self.respect_pc:
455 pc = self.pc.CIA.value
456 else:
457 pc = self.fake_pc
458 self._pc = pc
459 ins = self.imem.ld(pc, 4, False, True)
460 if ins is None:
461 raise KeyError("no instruction at 0x%x" % pc)
462 print("setup: 0x%x 0x%x %s" % (pc, ins & 0xffffffff, bin(ins)))
463 print ("NIA, CIA", self.pc.CIA.value, self.pc.NIA.value)
464
465 yield self.dec2.dec.raw_opcode_in.eq(ins & 0xffffffff)
466 yield self.dec2.dec.bigendian.eq(0) # little / big?
467
468 def execute_one(self):
469 """execute one instruction
470 """
471 # get the disassembly code for this instruction
472 code = self.disassembly[self._pc]
473 print("sim-execute", hex(self._pc), code)
474 opname = code.split(' ')[0]
475 yield from self.call(opname)
476
477 if not self.respect_pc:
478 self.fake_pc += 4
479 print ("NIA, CIA", self.pc.CIA.value, self.pc.NIA.value)
480
481 def get_assembly_name(self):
482 # TODO, asmregs is from the spec, e.g. add RT,RA,RB
483 # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
484 asmcode = yield self.dec2.dec.op.asmcode
485 asmop = insns.get(asmcode, None)
486
487 # sigh reconstruct the assembly instruction name
488 ov_en = yield self.dec2.e.oe.oe
489 ov_ok = yield self.dec2.e.oe.ok
490 if ov_en & ov_ok:
491 asmop += "."
492 lk = yield self.dec2.e.lk
493 if lk:
494 asmop += "l"
495 int_op = yield self.dec2.dec.op.internal_op
496 print ("int_op", int_op)
497 if int_op in [InternalOp.OP_B.value, InternalOp.OP_BC.value]:
498 AA = yield self.dec2.dec.fields.FormI.AA[0:-1]
499 print ("AA", AA)
500 if AA:
501 asmop += "a"
502 if int_op == InternalOp.OP_MFCR.value:
503 dec_insn = yield self.dec2.e.insn
504 if dec_insn & (1<<20) != 0: # sigh
505 asmop = 'mfocrf'
506 else:
507 asmop = 'mfcr'
508 # XXX TODO: for whatever weird reason this doesn't work
509 # https://bugs.libre-soc.org/show_bug.cgi?id=390
510 if int_op == InternalOp.OP_MTCRF.value:
511 dec_insn = yield self.dec2.e.insn
512 if dec_insn & (1<<20) != 0: # sigh
513 asmop = 'mtocrf'
514 else:
515 asmop = 'mtcrf'
516 return asmop
517
518 def call(self, name):
519 # TODO, asmregs is from the spec, e.g. add RT,RA,RB
520 # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
521 asmop = yield from self.get_assembly_name()
522 print ("call", name, asmop)
523 if name not in ['mtcrf', 'mtocrf']:
524 assert name == asmop, "name %s != %s" % (name, asmop)
525
526 info = self.instrs[name]
527 yield from self.prep_namespace(info.form, info.op_fields)
528
529 # preserve order of register names
530 input_names = create_args(list(info.read_regs) + list(info.uninit_regs))
531 print(input_names)
532
533 # main registers (RT, RA ...)
534 inputs = []
535 for name in input_names:
536 regnum = yield getattr(self.decoder, name)
537 regname = "_" + name
538 self.namespace[regname] = regnum
539 print('reading reg %d' % regnum)
540 inputs.append(self.gpr(regnum))
541
542 # "special" registers
543 for special in info.special_regs:
544 if special in special_sprs:
545 inputs.append(self.spr[special])
546 else:
547 inputs.append(self.namespace[special])
548
549 print(inputs)
550 results = info.func(self, *inputs)
551 print(results)
552
553 # detect if CA/CA32 already in outputs (sra*, basically)
554 already_done = 0
555 if info.write_regs:
556 output_names = create_args(info.write_regs)
557 for name in output_names:
558 if name == 'CA':
559 already_done |= 1
560 if name == 'CA32':
561 already_done |= 2
562
563 print ("carry already done?", bin(already_done))
564 carry_en = yield self.dec2.e.output_carry
565 if carry_en:
566 yield from self.handle_carry_(inputs, results, already_done)
567
568 # detect if overflow was in return result
569 overflow = None
570 if info.write_regs:
571 for name, output in zip(output_names, results):
572 if name == 'overflow':
573 overflow = output
574
575 ov_en = yield self.dec2.e.oe.oe
576 ov_ok = yield self.dec2.e.oe.ok
577 print ("internal overflow", overflow)
578 if ov_en & ov_ok:
579 yield from self.handle_overflow(inputs, results, overflow)
580
581 rc_en = yield self.dec2.e.rc.data
582 if rc_en:
583 self.handle_comparison(results)
584
585 # any modified return results?
586 if info.write_regs:
587 for name, output in zip(output_names, results):
588 if name == 'overflow': # ignore, done already (above)
589 continue
590 if isinstance(output, int):
591 output = SelectableInt(output, 256)
592 if name in ['CA', 'CA32']:
593 if carry_en:
594 print ("writing %s to XER" % name, output)
595 self.spr['XER'][XER_bits[name]] = output.value
596 else:
597 print ("NOT writing %s to XER" % name, output)
598 elif name in info.special_regs:
599 print('writing special %s' % name, output, special_sprs)
600 if name in special_sprs:
601 self.spr[name] = output
602 else:
603 self.namespace[name].eq(output)
604 if name == 'MSR':
605 print ('msr written', hex(self.msr.value))
606 else:
607 regnum = yield getattr(self.decoder, name)
608 print('writing reg %d %s' % (regnum, str(output)))
609 if output.bits > 64:
610 output = SelectableInt(output.value, 64)
611 self.gpr[regnum] = output
612
613 # update program counter
614 self.pc.update(self.namespace)
615
616
617 def inject():
618 """Decorator factory.
619
620 this decorator will "inject" variables into the function's namespace,
621 from the *dictionary* in self.namespace. it therefore becomes possible
622 to make it look like a whole stack of variables which would otherwise
623 need "self." inserted in front of them (*and* for those variables to be
624 added to the instance) "appear" in the function.
625
626 "self.namespace['SI']" for example becomes accessible as just "SI" but
627 *only* inside the function, when decorated.
628 """
629 def variable_injector(func):
630 @wraps(func)
631 def decorator(*args, **kwargs):
632 try:
633 func_globals = func.__globals__ # Python 2.6+
634 except AttributeError:
635 func_globals = func.func_globals # Earlier versions.
636
637 context = args[0].namespace # variables to be injected
638 saved_values = func_globals.copy() # Shallow copy of dict.
639 func_globals.update(context)
640 result = func(*args, **kwargs)
641 args[0].namespace = func_globals
642 #exec (func.__code__, func_globals)
643
644 #finally:
645 # func_globals = saved_values # Undo changes.
646
647 return result
648
649 return decorator
650
651 return variable_injector
652