1 # SPDX-License-Identifier: LGPLv3+
2 # Copyright (C) 2020, 2021 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
3 # Copyright (C) 2020 Michael Nolan
4 # Funded by NLnet http://nlnet.nl
5 """core of the python-based POWER9 simulator
7 this is part of a cycle-accurate POWER9 simulator. its primary purpose is
8 not speed, it is for both learning and educational purposes, as well as
9 a method of verifying the HDL.
13 * https://bugs.libre-soc.org/show_bug.cgi?id=424
16 from collections
import namedtuple
17 from copy
import deepcopy
18 from functools
import wraps
22 from openpower
.syscalls
import ppc_flags
24 from elftools
.elf
.elffile
import ELFFile
# for isinstance
26 from nmigen
.sim
import Settle
27 import openpower
.syscalls
28 from openpower
.consts
import (MSRb
, PIb
, # big-endian (PowerISA versions)
29 SVP64CROffs
, SVP64MODEb
)
30 from openpower
.decoder
.helpers
import (ISACallerHelper
, ISAFPHelpers
, exts
,
31 gtu
, undefined
, copy_assign_rhs
)
32 from openpower
.decoder
.isa
.mem
import Mem
, MemMMap
, MemException
, LoadedELF
33 from openpower
.decoder
.isa
.radixmmu
import RADIX
34 from openpower
.decoder
.isa
.svshape
import SVSHAPE
35 from openpower
.decoder
.isa
.svstate
import SVP64State
36 from openpower
.decoder
.orderedset
import OrderedSet
37 from openpower
.decoder
.power_enums
import (FPTRANS_INSNS
, CRInSel
, CROutSel
,
38 In1Sel
, In2Sel
, In3Sel
, LDSTMode
,
39 MicrOp
, OutSel
, SVMode
,
40 SVP64LDSTmode
, SVP64PredCR
,
41 SVP64PredInt
, SVP64PredMode
,
42 SVP64RMMode
, SVPType
, XER_bits
,
43 insns
, spr_byname
, spr_dict
,
45 from openpower
.insndb
.core
import SVP64Instruction
46 from openpower
.decoder
.power_svp64
import SVP64RM
, decode_extra
47 from openpower
.decoder
.selectable_int
import (FieldSelectableInt
,
48 SelectableInt
, selectconcat
,
49 EFFECTIVELY_UNLIMITED
)
50 from openpower
.consts
import DEFAULT_MSR
51 from openpower
.fpscr
import FPSCRState
52 from openpower
.xer
import XERState
53 from openpower
.util
import LogType
, log
55 LDST_UPDATE_INSNS
= ['ldu', 'lwzu', 'lbzu', 'lhzu', 'lhau', 'lfsu', 'lfdu',
56 'stwu', 'stbu', 'sthu', 'stfsu', 'stfdu', 'stdu',
60 instruction_info
= namedtuple('instruction_info',
61 'func read_regs uninit_regs write_regs ' +
62 'special_regs op_fields form asmregs')
72 # rrright. this is here basically because the compiler pywriter returns
73 # results in a specific priority order. to make sure regs match up they
74 # need partial sorting. sigh.
76 # TODO (lkcl): adjust other registers that should be in a particular order
77 # probably CA, CA32, and CR
105 "overflow": 7, # should definitely be last
109 fregs
= ['FRA', 'FRB', 'FRC', 'FRS', 'FRT']
112 def get_masked_reg(regs
, base
, offs
, ew_bits
):
113 # rrrright. start by breaking down into row/col, based on elwidth
114 gpr_offs
= offs
// (64 // ew_bits
)
115 gpr_col
= offs
% (64 // ew_bits
)
116 # compute the mask based on ew_bits
117 mask
= (1 << ew_bits
) - 1
118 # now select the 64-bit register, but get its value (easier)
119 val
= regs
[base
+ gpr_offs
]
120 # shift down so element we want is at LSB
121 val
>>= gpr_col
* ew_bits
122 # mask so we only return the LSB element
126 def set_masked_reg(regs
, base
, offs
, ew_bits
, value
):
127 # rrrright. start by breaking down into row/col, based on elwidth
128 gpr_offs
= offs
// (64//ew_bits
)
129 gpr_col
= offs
% (64//ew_bits
)
130 # compute the mask based on ew_bits
131 mask
= (1 << ew_bits
)-1
132 # now select the 64-bit register, but get its value (easier)
133 val
= regs
[base
+gpr_offs
]
134 # now mask out the bit we don't want
135 val
= val
& ~
(mask
<< (gpr_col
*ew_bits
))
136 # then wipe the bit we don't want from the value
138 # OR the new value in, shifted up
139 val |
= value
<< (gpr_col
*ew_bits
)
140 regs
[base
+gpr_offs
] = val
143 def create_args(reglist
, extra
=None):
144 retval
= list(OrderedSet(reglist
))
145 retval
.sort(key
=lambda reg
: REG_SORT_ORDER
.get(reg
, 0))
146 if extra
is not None:
147 return [extra
] + retval
151 def create_full_args(*, read_regs
, special_regs
, uninit_regs
, write_regs
,
154 *read_regs
, *uninit_regs
, *write_regs
, *special_regs
], extra
=extra
)
157 def is_ffirst_mode(dec2
):
158 rm_mode
= yield dec2
.rm_dec
.mode
159 return rm_mode
== SVP64RMMode
.FFIRST
.value
163 def __init__(self
, decoder
, isacaller
, svstate
, regfile
):
166 self
.isacaller
= isacaller
167 self
.svstate
= svstate
168 for i
in range(len(regfile
)):
169 self
[i
] = SelectableInt(regfile
[i
], 64)
171 def __call__(self
, ridx
, is_vec
=False, offs
=0, elwidth
=64):
172 if isinstance(ridx
, SelectableInt
):
175 return self
[ridx
+offs
]
176 # rrrright. start by breaking down into row/col, based on elwidth
177 gpr_offs
= offs
// (64//elwidth
)
178 gpr_col
= offs
% (64//elwidth
)
179 # now select the 64-bit register, but get its value (easier)
180 val
= self
[ridx
+gpr_offs
].value
181 # now shift down and mask out
182 val
= val
>> (gpr_col
*elwidth
) & ((1 << elwidth
)-1)
183 # finally, return a SelectableInt at the required elwidth
184 log("GPR call", ridx
, "isvec", is_vec
, "offs", offs
,
185 "elwid", elwidth
, "offs/col", gpr_offs
, gpr_col
, "val", hex(val
))
186 return SelectableInt(val
, elwidth
)
188 def set_form(self
, form
):
191 def write(self
, rnum
, value
, is_vec
=False, elwidth
=64):
193 if isinstance(rnum
, SelectableInt
):
195 if isinstance(value
, SelectableInt
):
198 if isinstance(rnum
, tuple):
199 rnum
, base
, offs
= rnum
202 # rrrright. start by breaking down into row/col, based on elwidth
203 gpr_offs
= offs
// (64//elwidth
)
204 gpr_col
= offs
% (64//elwidth
)
205 # compute the mask based on elwidth
206 mask
= (1 << elwidth
)-1
207 # now select the 64-bit register, but get its value (easier)
208 val
= self
[base
+gpr_offs
].value
209 # now mask out the bit we don't want
210 val
= val
& ~
(mask
<< (gpr_col
*elwidth
))
211 # then wipe the bit we don't want from the value
213 # OR the new value in, shifted up
214 val |
= value
<< (gpr_col
*elwidth
)
215 # finally put the damn value into the regfile
216 log("GPR write", base
, "isvec", is_vec
, "offs", offs
,
217 "elwid", elwidth
, "offs/col", gpr_offs
, gpr_col
, "val", hex(val
),
219 dict.__setitem
__(self
, base
+gpr_offs
, SelectableInt(val
, 64))
221 def __setitem__(self
, rnum
, value
):
222 # rnum = rnum.value # only SelectableInt allowed
223 log("GPR setitem", rnum
, value
)
224 if isinstance(rnum
, SelectableInt
):
226 dict.__setitem
__(self
, rnum
, value
)
228 def getz(self
, rnum
, rvalue
=None):
229 # rnum = rnum.value # only SelectableInt allowed
230 log("GPR getzero?", rnum
, rvalue
)
231 if rvalue
is not None:
233 return SelectableInt(0, rvalue
.bits
)
236 return SelectableInt(0, 64)
239 def _get_regnum(self
, attr
):
240 getform
= self
.sd
.sigforms
[self
.form
]
241 rnum
= getattr(getform
, attr
)
244 def ___getitem__(self
, attr
):
245 """ XXX currently not used
247 rnum
= self
._get
_regnum
(attr
)
248 log("GPR getitem", attr
, rnum
)
249 return self
.regfile
[rnum
]
251 def dump(self
, printout
=True, heading
="reg"):
253 for i
in range(len(self
)):
254 res
.append(self
[i
].value
)
256 for i
in range(0, len(res
), 8):
259 s
.append("%08x" % res
[i
+j
])
261 log(heading
, "%2d" % i
, s
, kind
=LogType
.InstrInOuts
)
266 def __init__(self
, dec2
, initial_sprs
={}, gpr
=None):
268 self
.gpr
= gpr
# for SVSHAPE[0-3]
270 for key
, v
in initial_sprs
.items():
271 if isinstance(key
, SelectableInt
):
273 key
= special_sprs
.get(key
, key
)
274 if isinstance(key
, int):
277 info
= spr_byname
[key
]
278 if not isinstance(v
, SelectableInt
):
279 v
= SelectableInt(v
, info
.length
)
282 def __getitem__(self
, key
):
284 #log("dict", self.items())
285 # if key in special_sprs get the special spr, otherwise return key
286 if isinstance(key
, SelectableInt
):
288 if isinstance(key
, int):
289 key
= spr_dict
[key
].SPR
290 key
= special_sprs
.get(key
, key
)
291 if key
== 'HSRR0': # HACK!
293 if key
== 'HSRR1': # HACK!
296 res
= dict.__getitem
__(self
, key
)
298 if isinstance(key
, int):
301 info
= spr_byname
[key
]
302 self
[key
] = SelectableInt(0, info
.length
)
303 res
= dict.__getitem
__(self
, key
)
304 #log("spr returning", key, res)
307 def __setitem__(self
, key
, value
):
308 if isinstance(key
, SelectableInt
):
310 if isinstance(key
, int):
311 key
= spr_dict
[key
].SPR
313 key
= special_sprs
.get(key
, key
)
314 if key
== 'HSRR0': # HACK!
315 self
.__setitem
__('SRR0', value
)
316 if key
== 'HSRR1': # HACK!
317 self
.__setitem
__('SRR1', value
)
319 value
= XERState(value
)
320 if key
in ('SVSHAPE0', 'SVSHAPE1', 'SVSHAPE2', 'SVSHAPE3'):
321 value
= SVSHAPE(value
, self
.gpr
)
322 log("setting spr", key
, value
)
323 dict.__setitem
__(self
, key
, value
)
325 def __call__(self
, ridx
):
328 def dump(self
, printout
=True):
330 keys
= list(self
.keys())
333 sprname
= spr_dict
.get(k
, None)
337 sprname
= sprname
.SPR
338 res
.append((sprname
, self
[k
].value
))
340 for sprname
, value
in res
:
341 print(" ", sprname
, hex(value
))
346 def __init__(self
, pc_init
=0):
347 self
.CIA
= SelectableInt(pc_init
, 64)
348 self
.NIA
= self
.CIA
+ SelectableInt(4, 64) # only true for v3.0B!
350 def update_nia(self
, is_svp64
):
351 increment
= 8 if is_svp64
else 4
352 self
.NIA
= self
.CIA
+ SelectableInt(increment
, 64)
354 def update(self
, namespace
, is_svp64
):
355 """updates the program counter (PC) by 4 if v3.0B mode or 8 if SVP64
357 self
.CIA
= namespace
['NIA'].narrow(64)
358 self
.update_nia(is_svp64
)
359 namespace
['CIA'] = self
.CIA
360 namespace
['NIA'] = self
.NIA
364 # See PowerISA Version 3.0 B Book 1
365 # Section 2.3.1 Condition Register pages 30 - 31
367 LT
= FL
= 0 # negative, less than, floating-point less than
368 GT
= FG
= 1 # positive, greater than, floating-point greater than
369 EQ
= FE
= 2 # equal, floating-point equal
370 SO
= FU
= 3 # summary overflow, floating-point unordered
372 def __init__(self
, init
=0):
373 # rev_cr = int('{:016b}'.format(initial_cr)[::-1], 2)
374 # self.cr = FieldSelectableInt(self._cr, list(range(32, 64)))
375 self
.cr
= SelectableInt(init
, 64) # underlying reg
376 # field-selectable versions of Condition Register TODO check bitranges?
379 bits
= tuple(range(i
*4+32, (i
+1)*4+32))
380 _cr
= FieldSelectableInt(self
.cr
, bits
)
384 # decode SVP64 predicate integer to reg number and invert
385 def get_predint(gpr
, mask
):
389 log("get_predint", mask
, SVP64PredInt
.ALWAYS
.value
)
390 if mask
== SVP64PredInt
.ALWAYS
.value
:
391 return 0xffff_ffff_ffff_ffff # 64 bits of 1
392 if mask
== SVP64PredInt
.R3_UNARY
.value
:
393 return 1 << (r3
.value
& 0b111111)
394 if mask
== SVP64PredInt
.R3
.value
:
396 if mask
== SVP64PredInt
.R3_N
.value
:
398 if mask
== SVP64PredInt
.R10
.value
:
400 if mask
== SVP64PredInt
.R10_N
.value
:
402 if mask
== SVP64PredInt
.R30
.value
:
404 if mask
== SVP64PredInt
.R30_N
.value
:
408 # decode SVP64 predicate CR to reg number and invert status
409 def _get_predcr(mask
):
410 if mask
== SVP64PredCR
.LT
.value
:
412 if mask
== SVP64PredCR
.GE
.value
:
414 if mask
== SVP64PredCR
.GT
.value
:
416 if mask
== SVP64PredCR
.LE
.value
:
418 if mask
== SVP64PredCR
.EQ
.value
:
420 if mask
== SVP64PredCR
.NE
.value
:
422 if mask
== SVP64PredCR
.SO
.value
:
424 if mask
== SVP64PredCR
.NS
.value
:
428 # read individual CR fields (0..VL-1), extract the required bit
429 # and construct the mask
430 def get_predcr(crl
, predselect
, vl
):
431 idx
, noninv
= _get_predcr(predselect
)
434 cr
= crl
[i
+SVP64CROffs
.CRPred
]
435 if cr
[idx
].value
== noninv
:
437 log("get_predcr", vl
, idx
, noninv
, i
+SVP64CROffs
.CRPred
,
438 bin(cr
.asint()), cr
[idx
].value
, bin(mask
))
442 # TODO, really should just be using PowerDecoder2
443 def get_idx_map(dec2
, name
):
445 in1_sel
= yield op
.in1_sel
446 in2_sel
= yield op
.in2_sel
447 in3_sel
= yield op
.in3_sel
448 in1
= yield dec2
.e
.read_reg1
.data
449 # identify which regnames map to in1/2/3
450 if name
== 'RA' or name
== 'RA_OR_ZERO':
451 if (in1_sel
== In1Sel
.RA
.value
or
452 (in1_sel
== In1Sel
.RA_OR_ZERO
.value
and in1
!= 0)):
454 if in1_sel
== In1Sel
.RA_OR_ZERO
.value
:
457 if in2_sel
== In2Sel
.RB
.value
:
459 if in3_sel
== In3Sel
.RB
.value
:
461 # XXX TODO, RC doesn't exist yet!
463 if in3_sel
== In3Sel
.RC
.value
:
465 elif name
in ['EA', 'RS']:
466 if in1_sel
== In1Sel
.RS
.value
:
468 if in2_sel
== In2Sel
.RS
.value
:
470 if in3_sel
== In3Sel
.RS
.value
:
473 if in1_sel
== In1Sel
.FRA
.value
:
475 if in3_sel
== In3Sel
.FRA
.value
:
478 if in2_sel
== In2Sel
.FRB
.value
:
481 if in3_sel
== In3Sel
.FRC
.value
:
484 if in1_sel
== In1Sel
.FRS
.value
:
486 if in3_sel
== In3Sel
.FRS
.value
:
489 if in1_sel
== In1Sel
.FRT
.value
:
492 if in1_sel
== In1Sel
.RT
.value
:
497 # TODO, really should just be using PowerDecoder2
498 def get_idx_in(dec2
, name
, ewmode
=False):
499 idx
= yield from get_idx_map(dec2
, name
)
503 in1_sel
= yield op
.in1_sel
504 in2_sel
= yield op
.in2_sel
505 in3_sel
= yield op
.in3_sel
506 # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
507 in1
= yield dec2
.e
.read_reg1
.data
508 in2
= yield dec2
.e
.read_reg2
.data
509 in3
= yield dec2
.e
.read_reg3
.data
511 in1_base
= yield dec2
.e
.read_reg1
.base
512 in2_base
= yield dec2
.e
.read_reg2
.base
513 in3_base
= yield dec2
.e
.read_reg3
.base
514 in1_offs
= yield dec2
.e
.read_reg1
.offs
515 in2_offs
= yield dec2
.e
.read_reg2
.offs
516 in3_offs
= yield dec2
.e
.read_reg3
.offs
517 in1
= (in1
, in1_base
, in1_offs
)
518 in2
= (in2
, in2_base
, in2_offs
)
519 in3
= (in3
, in3_base
, in3_offs
)
521 in1_isvec
= yield dec2
.in1_isvec
522 in2_isvec
= yield dec2
.in2_isvec
523 in3_isvec
= yield dec2
.in3_isvec
524 log("get_idx_in in1", name
, in1_sel
, In1Sel
.RA
.value
,
526 log("get_idx_in in2", name
, in2_sel
, In2Sel
.RB
.value
,
528 log("get_idx_in in3", name
, in3_sel
, In3Sel
.RS
.value
,
530 log("get_idx_in FRS in3", name
, in3_sel
, In3Sel
.FRS
.value
,
532 log("get_idx_in FRB in2", name
, in2_sel
, In2Sel
.FRB
.value
,
534 log("get_idx_in FRC in3", name
, in3_sel
, In3Sel
.FRC
.value
,
537 return in1
, in1_isvec
539 return in2
, in2_isvec
541 return in3
, in3_isvec
545 # TODO, really should just be using PowerDecoder2
546 def get_cr_in(dec2
, name
):
548 in_sel
= yield op
.cr_in
549 in_bitfield
= yield dec2
.dec_cr_in
.cr_bitfield
.data
550 sv_cr_in
= yield op
.sv_cr_in
551 spec
= yield dec2
.crin_svdec
.spec
552 sv_override
= yield dec2
.dec_cr_in
.sv_override
553 # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
554 in1
= yield dec2
.e
.read_cr1
.data
555 cr_isvec
= yield dec2
.cr_in_isvec
556 log("get_cr_in", in_sel
, CROutSel
.CR0
.value
, in1
, cr_isvec
)
557 log(" sv_cr_in", sv_cr_in
)
558 log(" cr_bf", in_bitfield
)
560 log(" override", sv_override
)
561 # identify which regnames map to in / o2
563 if in_sel
== CRInSel
.BI
.value
:
565 if name
in ['BA', 'BB']:
566 if in_sel
== CRInSel
.BA_BB
.value
:
569 if in_sel
== CRInSel
.BFA
.value
:
571 log("get_cr_in not found", name
)
575 # TODO, really should just be using PowerDecoder2
576 def get_cr_out(dec2
, name
):
578 out_sel
= yield op
.cr_out
579 out_bitfield
= yield dec2
.dec_cr_out
.cr_bitfield
.data
580 sv_cr_out
= yield op
.sv_cr_out
581 spec
= yield dec2
.crout_svdec
.spec
582 sv_override
= yield dec2
.dec_cr_out
.sv_override
583 # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
584 out
= yield dec2
.e
.write_cr
.data
585 o_isvec
= yield dec2
.cr_out_isvec
586 log("get_cr_out", out_sel
, CROutSel
.CR0
.value
, out
, o_isvec
)
587 log(" sv_cr_out", sv_cr_out
)
588 log(" cr_bf", out_bitfield
)
590 log(" override", sv_override
)
591 # identify which regnames map to out / o2
593 if out_sel
== CROutSel
.BF
.value
:
596 if out_sel
== CROutSel
.BT
.value
:
599 if out_sel
== CROutSel
.CR0
.value
:
601 if name
== 'CR1': # these are not actually calculated correctly
602 if out_sel
== CROutSel
.CR1
.value
:
604 # check RC1 set? if so return implicit vector, this is a REAL bad hack
605 RC1
= yield dec2
.rm_dec
.RC1
607 log("get_cr_out RC1 mode")
609 return 0, True # XXX TODO: offset CR0 from SVSTATE SPR
611 return 1, True # XXX TODO: offset CR1 from SVSTATE SPR
613 log("get_cr_out not found", name
)
617 # TODO, really should just be using PowerDecoder2
618 def get_out_map(dec2
, name
):
620 out_sel
= yield op
.out_sel
621 # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
622 out
= yield dec2
.e
.write_reg
.data
623 # identify which regnames map to out / o2
625 if out_sel
== OutSel
.RA
.value
:
628 if out_sel
== OutSel
.RT
.value
:
630 if out_sel
== OutSel
.RT_OR_ZERO
.value
and out
!= 0:
632 elif name
== 'RT_OR_ZERO':
633 if out_sel
== OutSel
.RT_OR_ZERO
.value
:
636 if out_sel
== OutSel
.FRA
.value
:
639 if out_sel
== OutSel
.FRS
.value
:
642 if out_sel
== OutSel
.FRT
.value
:
647 # TODO, really should just be using PowerDecoder2
648 def get_idx_out(dec2
, name
, ewmode
=False):
650 out_sel
= yield op
.out_sel
651 # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
652 out
= yield dec2
.e
.write_reg
.data
653 o_isvec
= yield dec2
.o_isvec
655 offs
= yield dec2
.e
.write_reg
.offs
656 base
= yield dec2
.e
.write_reg
.base
657 out
= (out
, base
, offs
)
658 # identify which regnames map to out / o2
659 ismap
= yield from get_out_map(dec2
, name
)
661 log("get_idx_out", name
, out_sel
, out
, o_isvec
)
663 log("get_idx_out not found", name
, out_sel
, out
, o_isvec
)
667 # TODO, really should just be using PowerDecoder2
668 def get_out2_map(dec2
, name
):
669 # check first if register is activated for write
671 out_sel
= yield op
.out_sel
672 out
= yield dec2
.e
.write_ea
.data
673 out_ok
= yield dec2
.e
.write_ea
.ok
677 if name
in ['EA', 'RA']:
678 if hasattr(op
, "upd"):
679 # update mode LD/ST uses read-reg A also as an output
681 log("get_idx_out2", upd
, LDSTMode
.update
.value
,
682 out_sel
, OutSel
.RA
.value
,
684 if upd
== LDSTMode
.update
.value
:
687 fft_en
= yield dec2
.implicit_rs
689 log("get_idx_out2", out_sel
, OutSel
.RS
.value
,
693 fft_en
= yield dec2
.implicit_rs
695 log("get_idx_out2", out_sel
, OutSel
.FRS
.value
,
701 # TODO, really should just be using PowerDecoder2
702 def get_idx_out2(dec2
, name
, ewmode
=False):
703 # check first if register is activated for write
705 out_sel
= yield op
.out_sel
706 out
= yield dec2
.e
.write_ea
.data
708 offs
= yield dec2
.e
.write_ea
.offs
709 base
= yield dec2
.e
.write_ea
.base
710 out
= (out
, base
, offs
)
711 o_isvec
= yield dec2
.o2_isvec
712 ismap
= yield from get_out2_map(dec2
, name
)
714 log("get_idx_out2", name
, out_sel
, out
, o_isvec
)
720 """deals with svstate looping.
723 def __init__(self
, svstate
):
724 self
.svstate
= svstate
727 def new_iterators(self
):
728 self
.src_it
= self
.src_iterator()
729 self
.dst_it
= self
.dst_iterator()
733 self
.new_ssubstep
= 0
734 self
.new_dsubstep
= 0
735 self
.pred_dst_zero
= 0
736 self
.pred_src_zero
= 0
738 def src_iterator(self
):
739 """source-stepping iterator
741 pack
= self
.svstate
.pack
745 # pack advances subvl in *outer* loop
746 while True: # outer subvl loop
747 while True: # inner vl loop
750 srcmask
= self
.srcmask
751 srcstep
= self
.svstate
.srcstep
752 pred_src_zero
= ((1 << srcstep
) & srcmask
) != 0
753 if self
.pred_sz
or pred_src_zero
:
754 self
.pred_src_zero
= not pred_src_zero
755 log(" advance src", srcstep
, vl
,
756 self
.svstate
.ssubstep
, subvl
)
757 # yield actual substep/srcstep
758 yield (self
.svstate
.ssubstep
, srcstep
)
759 # the way yield works these could have been modified.
762 srcstep
= self
.svstate
.srcstep
763 log(" advance src check", srcstep
, vl
,
764 self
.svstate
.ssubstep
, subvl
, srcstep
== vl
-1,
765 self
.svstate
.ssubstep
== subvl
)
766 if srcstep
== vl
-1: # end-point
767 self
.svstate
.srcstep
= SelectableInt(0, 7) # reset
768 if self
.svstate
.ssubstep
== subvl
: # end-point
769 log(" advance pack stop")
771 break # exit inner loop
772 self
.svstate
.srcstep
+= SelectableInt(1, 7) # advance ss
774 if self
.svstate
.ssubstep
== subvl
: # end-point
775 self
.svstate
.ssubstep
= SelectableInt(0, 2) # reset
776 log(" advance pack stop")
778 self
.svstate
.ssubstep
+= SelectableInt(1, 2)
781 # these cannot be done as for-loops because SVSTATE may change
782 # (srcstep/substep may be modified, interrupted, subvl/vl change)
783 # but they *can* be done as while-loops as long as every SVSTATE
784 # "thing" is re-read every single time a yield gives indices
785 while True: # outer vl loop
786 while True: # inner subvl loop
789 srcmask
= self
.srcmask
790 srcstep
= self
.svstate
.srcstep
791 pred_src_zero
= ((1 << srcstep
) & srcmask
) != 0
792 if self
.pred_sz
or pred_src_zero
:
793 self
.pred_src_zero
= not pred_src_zero
794 log(" advance src", srcstep
, vl
,
795 self
.svstate
.ssubstep
, subvl
)
796 # yield actual substep/srcstep
797 yield (self
.svstate
.ssubstep
, srcstep
)
798 if self
.svstate
.ssubstep
== subvl
: # end-point
799 self
.svstate
.ssubstep
= SelectableInt(0, 2) # reset
800 break # exit inner loop
801 self
.svstate
.ssubstep
+= SelectableInt(1, 2)
803 if srcstep
== vl
-1: # end-point
804 self
.svstate
.srcstep
= SelectableInt(0, 7) # reset
807 self
.svstate
.srcstep
+= SelectableInt(1, 7) # advance srcstep
809 def dst_iterator(self
):
810 """dest-stepping iterator
812 unpack
= self
.svstate
.unpack
816 # pack advances subvl in *outer* loop
817 while True: # outer subvl loop
818 while True: # inner vl loop
821 dstmask
= self
.dstmask
822 dststep
= self
.svstate
.dststep
823 pred_dst_zero
= ((1 << dststep
) & dstmask
) != 0
824 if self
.pred_dz
or pred_dst_zero
:
825 self
.pred_dst_zero
= not pred_dst_zero
826 log(" advance dst", dststep
, vl
,
827 self
.svstate
.dsubstep
, subvl
)
828 # yield actual substep/dststep
829 yield (self
.svstate
.dsubstep
, dststep
)
830 # the way yield works these could have been modified.
832 dststep
= self
.svstate
.dststep
833 log(" advance dst check", dststep
, vl
,
834 self
.svstate
.ssubstep
, subvl
)
835 if dststep
== vl
-1: # end-point
836 self
.svstate
.dststep
= SelectableInt(0, 7) # reset
837 if self
.svstate
.dsubstep
== subvl
: # end-point
838 log(" advance unpack stop")
841 self
.svstate
.dststep
+= SelectableInt(1, 7) # advance ds
843 if self
.svstate
.dsubstep
== subvl
: # end-point
844 self
.svstate
.dsubstep
= SelectableInt(0, 2) # reset
845 log(" advance unpack stop")
847 self
.svstate
.dsubstep
+= SelectableInt(1, 2)
849 # these cannot be done as for-loops because SVSTATE may change
850 # (dststep/substep may be modified, interrupted, subvl/vl change)
851 # but they *can* be done as while-loops as long as every SVSTATE
852 # "thing" is re-read every single time a yield gives indices
853 while True: # outer vl loop
854 while True: # inner subvl loop
856 dstmask
= self
.dstmask
857 dststep
= self
.svstate
.dststep
858 pred_dst_zero
= ((1 << dststep
) & dstmask
) != 0
859 if self
.pred_dz
or pred_dst_zero
:
860 self
.pred_dst_zero
= not pred_dst_zero
861 log(" advance dst", dststep
, self
.svstate
.vl
,
862 self
.svstate
.dsubstep
, subvl
)
863 # yield actual substep/dststep
864 yield (self
.svstate
.dsubstep
, dststep
)
865 if self
.svstate
.dsubstep
== subvl
: # end-point
866 self
.svstate
.dsubstep
= SelectableInt(0, 2) # reset
868 self
.svstate
.dsubstep
+= SelectableInt(1, 2)
871 if dststep
== vl
-1: # end-point
872 self
.svstate
.dststep
= SelectableInt(0, 7) # reset
874 self
.svstate
.dststep
+= SelectableInt(1, 7) # advance dststep
876 def src_iterate(self
):
877 """source-stepping iterator
881 pack
= self
.svstate
.pack
882 unpack
= self
.svstate
.unpack
883 ssubstep
= self
.svstate
.ssubstep
884 end_ssub
= ssubstep
== subvl
885 end_src
= self
.svstate
.srcstep
== vl
-1
886 log(" pack/unpack/subvl", pack
, unpack
, subvl
,
890 srcstep
= self
.svstate
.srcstep
891 srcmask
= self
.srcmask
893 # pack advances subvl in *outer* loop
895 assert srcstep
<= vl
-1
896 end_src
= srcstep
== vl
-1
901 self
.svstate
.ssubstep
+= SelectableInt(1, 2)
905 srcstep
+= 1 # advance srcstep
906 if not self
.srcstep_skip
:
908 if ((1 << srcstep
) & srcmask
) != 0:
911 log(" sskip", bin(srcmask
), bin(1 << srcstep
))
913 # advance subvl in *inner* loop
916 assert srcstep
<= vl
-1
917 end_src
= srcstep
== vl
-1
918 if end_src
: # end-point
924 if not self
.srcstep_skip
:
926 if ((1 << srcstep
) & srcmask
) != 0:
929 log(" sskip", bin(srcmask
), bin(1 << srcstep
))
930 self
.svstate
.ssubstep
= SelectableInt(0, 2) # reset
933 self
.svstate
.ssubstep
+= SelectableInt(1, 2)
935 self
.svstate
.srcstep
= SelectableInt(srcstep
, 7)
936 log(" advance src", self
.svstate
.srcstep
, self
.svstate
.ssubstep
,
939 def dst_iterate(self
):
940 """dest step iterator
944 pack
= self
.svstate
.pack
945 unpack
= self
.svstate
.unpack
946 dsubstep
= self
.svstate
.dsubstep
947 end_dsub
= dsubstep
== subvl
948 dststep
= self
.svstate
.dststep
949 end_dst
= dststep
== vl
-1
950 dstmask
= self
.dstmask
951 log(" pack/unpack/subvl", pack
, unpack
, subvl
,
956 # unpack advances subvl in *outer* loop
958 assert dststep
<= vl
-1
959 end_dst
= dststep
== vl
-1
964 self
.svstate
.dsubstep
+= SelectableInt(1, 2)
968 dststep
+= 1 # advance dststep
969 if not self
.dststep_skip
:
971 if ((1 << dststep
) & dstmask
) != 0:
974 log(" dskip", bin(dstmask
), bin(1 << dststep
))
976 # advance subvl in *inner* loop
979 assert dststep
<= vl
-1
980 end_dst
= dststep
== vl
-1
981 if end_dst
: # end-point
987 if not self
.dststep_skip
:
989 if ((1 << dststep
) & dstmask
) != 0:
992 log(" dskip", bin(dstmask
), bin(1 << dststep
))
993 self
.svstate
.dsubstep
= SelectableInt(0, 2) # reset
996 self
.svstate
.dsubstep
+= SelectableInt(1, 2)
998 self
.svstate
.dststep
= SelectableInt(dststep
, 7)
999 log(" advance dst", self
.svstate
.dststep
, self
.svstate
.dsubstep
,
1002 def at_loopend(self
):
1003 """tells if this is the last possible element. uses the cached values
1004 for src/dst-step and sub-steps
1007 vl
= self
.svstate
.vl
1008 srcstep
, dststep
= self
.new_srcstep
, self
.new_dststep
1009 ssubstep
, dsubstep
= self
.new_ssubstep
, self
.new_dsubstep
1010 end_ssub
= ssubstep
== subvl
1011 end_dsub
= dsubstep
== subvl
1012 if srcstep
== vl
-1 and end_ssub
:
1014 if dststep
== vl
-1 and end_dsub
:
1018 def advance_svstate_steps(self
):
1019 """ advance sub/steps. note that Pack/Unpack *INVERTS* the order.
1020 TODO when Pack/Unpack is set, substep becomes the *outer* loop
1022 self
.subvl
= yield self
.dec2
.rm_dec
.rm_in
.subvl
1023 if self
.loopend
: # huhn??
1028 def read_src_mask(self
):
1029 """read/update pred_sz and src mask
1031 # get SVSTATE VL (oh and print out some debug stuff)
1032 vl
= self
.svstate
.vl
1033 srcstep
= self
.svstate
.srcstep
1034 ssubstep
= self
.svstate
.ssubstep
1036 # get predicate mask (all 64 bits)
1037 srcmask
= 0xffff_ffff_ffff_ffff
1039 pmode
= yield self
.dec2
.rm_dec
.predmode
1040 sv_ptype
= yield self
.dec2
.dec
.op
.SV_Ptype
1041 srcpred
= yield self
.dec2
.rm_dec
.srcpred
1042 dstpred
= yield self
.dec2
.rm_dec
.dstpred
1043 pred_sz
= yield self
.dec2
.rm_dec
.pred_sz
1044 if pmode
== SVP64PredMode
.INT
.value
:
1045 srcmask
= dstmask
= get_predint(self
.gpr
, dstpred
)
1046 if sv_ptype
== SVPType
.P2
.value
:
1047 srcmask
= get_predint(self
.gpr
, srcpred
)
1048 elif pmode
== SVP64PredMode
.CR
.value
:
1049 srcmask
= dstmask
= get_predcr(self
.crl
, dstpred
, vl
)
1050 if sv_ptype
== SVPType
.P2
.value
:
1051 srcmask
= get_predcr(self
.crl
, srcpred
, vl
)
1052 # work out if the ssubsteps are completed
1053 ssubstart
= ssubstep
== 0
1054 log(" pmode", pmode
)
1055 log(" ptype", sv_ptype
)
1056 log(" srcpred", bin(srcpred
))
1057 log(" srcmask", bin(srcmask
))
1058 log(" pred_sz", bin(pred_sz
))
1059 log(" ssubstart", ssubstart
)
1061 # store all that above
1062 self
.srcstep_skip
= False
1063 self
.srcmask
= srcmask
1064 self
.pred_sz
= pred_sz
1065 self
.new_ssubstep
= ssubstep
1066 log(" new ssubstep", ssubstep
)
1067 # until the predicate mask has a "1" bit... or we run out of VL
1068 # let srcstep==VL be the indicator to move to next instruction
1070 self
.srcstep_skip
= True
1072 def read_dst_mask(self
):
1073 """same as read_src_mask - check and record everything needed
1075 # get SVSTATE VL (oh and print out some debug stuff)
1076 # yield Delay(1e-10) # make changes visible
1077 vl
= self
.svstate
.vl
1078 dststep
= self
.svstate
.dststep
1079 dsubstep
= self
.svstate
.dsubstep
1081 # get predicate mask (all 64 bits)
1082 dstmask
= 0xffff_ffff_ffff_ffff
1084 pmode
= yield self
.dec2
.rm_dec
.predmode
1085 reverse_gear
= yield self
.dec2
.rm_dec
.reverse_gear
1086 sv_ptype
= yield self
.dec2
.dec
.op
.SV_Ptype
1087 dstpred
= yield self
.dec2
.rm_dec
.dstpred
1088 pred_dz
= yield self
.dec2
.rm_dec
.pred_dz
1089 if pmode
== SVP64PredMode
.INT
.value
:
1090 dstmask
= get_predint(self
.gpr
, dstpred
)
1091 elif pmode
== SVP64PredMode
.CR
.value
:
1092 dstmask
= get_predcr(self
.crl
, dstpred
, vl
)
1093 # work out if the ssubsteps are completed
1094 dsubstart
= dsubstep
== 0
1095 log(" pmode", pmode
)
1096 log(" ptype", sv_ptype
)
1097 log(" dstpred", bin(dstpred
))
1098 log(" dstmask", bin(dstmask
))
1099 log(" pred_dz", bin(pred_dz
))
1100 log(" dsubstart", dsubstart
)
1102 self
.dststep_skip
= False
1103 self
.dstmask
= dstmask
1104 self
.pred_dz
= pred_dz
1105 self
.new_dsubstep
= dsubstep
1106 log(" new dsubstep", dsubstep
)
1108 self
.dststep_skip
= True
1110 def svstate_pre_inc(self
):
1111 """check if srcstep/dststep need to skip over masked-out predicate bits
1112 note that this is not supposed to do anything to substep,
1113 it is purely for skipping masked-out bits
1116 self
.subvl
= yield self
.dec2
.rm_dec
.rm_in
.subvl
1117 yield from self
.read_src_mask()
1118 yield from self
.read_dst_mask()
1125 srcstep
= self
.svstate
.srcstep
1126 srcmask
= self
.srcmask
1127 pred_src_zero
= self
.pred_sz
1128 vl
= self
.svstate
.vl
1129 # srcstep-skipping opportunity identified
1130 if self
.srcstep_skip
:
1131 # cannot do this with sv.bc - XXX TODO
1134 while (((1 << srcstep
) & srcmask
) == 0) and (srcstep
!= vl
):
1135 log(" sskip", bin(1 << srcstep
))
1138 # now work out if the relevant mask bits require zeroing
1140 pred_src_zero
= ((1 << srcstep
) & srcmask
) == 0
1142 # store new srcstep / dststep
1143 self
.new_srcstep
= srcstep
1144 self
.pred_src_zero
= pred_src_zero
1145 log(" new srcstep", srcstep
)
1148 # dststep-skipping opportunity identified
1149 dststep
= self
.svstate
.dststep
1150 dstmask
= self
.dstmask
1151 pred_dst_zero
= self
.pred_dz
1152 vl
= self
.svstate
.vl
1153 if self
.dststep_skip
:
1154 # cannot do this with sv.bc - XXX TODO
1157 while (((1 << dststep
) & dstmask
) == 0) and (dststep
!= vl
):
1158 log(" dskip", bin(1 << dststep
))
1161 # now work out if the relevant mask bits require zeroing
1163 pred_dst_zero
= ((1 << dststep
) & dstmask
) == 0
1165 # store new srcstep / dststep
1166 self
.new_dststep
= dststep
1167 self
.pred_dst_zero
= pred_dst_zero
1168 log(" new dststep", dststep
)
1171 class ExitSyscallCalled(Exception):
1175 class SyscallEmulator(openpower
.syscalls
.Dispatcher
):
1176 def __init__(self
, isacaller
):
1177 self
.__isacaller
= isacaller
1179 host
= os
.uname().machine
1180 bits
= (64 if (sys
.maxsize
> (2**32)) else 32)
1181 host
= openpower
.syscalls
.architecture(arch
=host
, bits
=bits
)
1183 return super().__init
__(guest
="ppc64", host
=host
)
1185 def __call__(self
, identifier
, *arguments
):
1186 (identifier
, *arguments
) = map(int, (identifier
, *arguments
))
1187 return super().__call
__(identifier
, *arguments
)
1189 def sys_exit_group(self
, status
, *rest
):
1190 self
.__isacaller
.halted
= True
1191 raise ExitSyscallCalled(status
)
1193 def sys_write(self
, fd
, buf
, count
, *rest
):
1195 buf
= self
.__isacaller
.mem
.get_ctypes(buf
, count
, is_write
=False)
1199 return os
.write(fd
, buf
)
1200 except OSError as e
:
1203 def sys_writev(self
, fd
, iov
, iovcnt
, *rest
):
1205 if iovcnt
< 0 or iovcnt
> IOV_MAX
:
1206 return -errno
.EINVAL
1207 struct_iovec
= struct
.Struct("<QQ")
1210 iov
= self
.__isacaller
.mem
.get_ctypes(
1211 iov
, struct_iovec
.size
* iovcnt
, is_write
=False)
1212 iov
= list(struct_iovec
.iter_unpack(iov
))
1215 for i
, iovec
in enumerate(iov
):
1216 iov_base
, iov_len
= iovec
1217 iov
[i
] = self
.__isacaller
.mem
.get_ctypes(
1218 iov_base
, iov_len
, is_write
=False)
1219 except (ValueError, MemException
):
1220 return -errno
.EFAULT
1222 return os
.writev(fd
, iov
)
1223 except OSError as e
:
1226 def sys_read(self
, fd
, buf
, count
, *rest
):
1228 buf
= self
.__isacaller
.mem
.get_ctypes(buf
, count
, is_write
=True)
1232 return os
.readv(fd
, [buf
])
1233 except OSError as e
:
1236 def sys_mmap(self
, addr
, length
, prot
, flags
, fd
, offset
, *rest
):
1237 return self
.__isacaller
.mem
.mmap_syscall(
1238 addr
, length
, prot
, flags
, fd
, offset
, is_mmap2
=False)
1240 def sys_mmap2(self
, addr
, length
, prot
, flags
, fd
, offset
, *rest
):
1241 return self
.__isacaller
.mem
.mmap_syscall(
1242 addr
, length
, prot
, flags
, fd
, offset
, is_mmap2
=True)
1244 def sys_brk(self
, addr
, *rest
):
1245 return self
.__isacaller
.mem
.brk_syscall(addr
)
1247 def sys_munmap(self
, addr
, length
, *rest
):
1248 return -errno
.ENOSYS
# TODO: implement
1250 def sys_mprotect(self
, addr
, length
, prot
, *rest
):
1251 return -errno
.ENOSYS
# TODO: implement
1253 def sys_pkey_mprotect(self
, addr
, length
, prot
, pkey
, *rest
):
1254 return -errno
.ENOSYS
# TODO: implement
1256 def sys_openat(self
, dirfd
, pathname
, flags
, mode
, *rest
):
1258 path
= self
.__isacaller
.mem
.read_cstr(pathname
)
1259 except (ValueError, MemException
):
1260 return -errno
.EFAULT
1262 if dirfd
== ppc_flags
.AT_FDCWD
:
1263 return os
.open(path
, flags
, mode
)
1265 return os
.open(path
, flags
, mode
, dir_fd
=dirfd
)
1266 except OSError as e
:
1272 nodename
= uname
.nodename
.encode()
1273 release
= b
'5.6.0-1-powerpc64le'
1274 version
= b
'#1 SMP Debian 5.6.7-1 (2020-04-29)'
1275 machine
= b
'ppc64le'
1277 return sysname
, nodename
, release
, version
, machine
, domainname
1279 def sys_uname(self
, buf
, *rest
):
1280 s
= struct
.Struct("<65s65s65s65s65s")
1282 buf
= self
.__isacaller
.mem
.get_ctypes(buf
, s
.size
, is_write
=True)
1283 except (ValueError, MemException
):
1284 return -errno
.EFAULT
1285 sysname
, nodename
, release
, version
, machine
, domainname
= \
1287 s
.pack_into(buf
, 0, sysname
, nodename
, release
, version
, machine
)
1290 def sys_newuname(self
, buf
, *rest
):
1291 name_len
= ppc_flags
.__NEW
_UTS
_LEN
+ 1
1292 s
= struct
.Struct("<%ds%ds%ds%ds%ds%ds" % ((name_len
,) * 6))
1294 buf
= self
.__isacaller
.mem
.get_ctypes(buf
, s
.size
, is_write
=True)
1295 except (ValueError, MemException
):
1296 return -errno
.EFAULT
1297 sysname
, nodename
, release
, version
, machine
, domainname
= \
1300 sysname
, nodename
, release
, version
, machine
, domainname
)
1303 def sys_readlink(self
, pathname
, buf
, bufsiz
, *rest
):
1304 dirfd
= ppc_flags
.AT_FDCWD
1305 return self
.sys_readlinkat(dirfd
, pathname
, buf
, bufsiz
)
1307 def sys_readlinkat(self
, dirfd
, pathname
, buf
, bufsiz
, *rest
):
1309 path
= self
.__isacaller
.mem
.read_cstr(pathname
)
1311 buf
= self
.__isacaller
.mem
.get_ctypes(
1312 buf
, bufsiz
, is_write
=True)
1315 except (ValueError, MemException
):
1316 return -errno
.EFAULT
1318 if dirfd
== ppc_flags
.AT_FDCWD
:
1319 result
= os
.readlink(path
)
1321 result
= os
.readlink(path
, dir_fd
=dirfd
)
1322 retval
= min(len(result
), len(buf
))
1323 buf
[:retval
] = result
[:retval
]
1325 except OSError as e
:
1329 class ISACaller(ISACallerHelper
, ISAFPHelpers
, StepLoop
):
1330 # decoder2 - an instance of power_decoder2
1331 # regfile - a list of initial values for the registers
1332 # initial_{etc} - initial values for SPRs, Condition Register, Mem, MSR
1333 # respect_pc - tracks the program counter. requires initial_insns
1334 def __init__(self
, decoder2
, regfile
, initial_sprs
=None, initial_cr
=0,
1335 initial_mem
=None, initial_msr
=0,
1348 use_syscall_emu
=False,
1349 emulating_mmap
=False,
1350 real_page_size
=None):
1352 self
.syscall
= SyscallEmulator(isacaller
=self
)
1353 if not use_mmap_mem
:
1354 log("forcing use_mmap_mem due to use_syscall_emu active")
1359 # we will eventually be able to load ELF files without use_syscall_emu
1360 # (e.g. the linux kernel), so do it in a separate if block
1361 if isinstance(initial_insns
, ELFFile
):
1362 if not use_mmap_mem
:
1363 log("forcing use_mmap_mem due to loading an ELF file")
1365 if not emulating_mmap
:
1366 log("forcing emulating_mmap due to loading an ELF file")
1367 emulating_mmap
= True
1369 # trace log file for model output. if None do nothing
1370 self
.insnlog
= insnlog
1371 self
.insnlog_is_file
= hasattr(insnlog
, "write")
1372 if not self
.insnlog_is_file
and self
.insnlog
:
1373 self
.insnlog
= open(self
.insnlog
, "w")
1375 self
.bigendian
= bigendian
1377 self
.is_svp64_mode
= False
1378 self
.respect_pc
= respect_pc
1379 if initial_sprs
is None:
1381 if initial_mem
is None:
1383 if fpregfile
is None:
1384 fpregfile
= [0] * 32
1385 if initial_insns
is None:
1387 assert self
.respect_pc
== False, "instructions required to honor pc"
1388 if initial_msr
is None:
1389 initial_msr
= DEFAULT_MSR
1391 log("ISACaller insns", respect_pc
, initial_insns
, disassembly
)
1392 log("ISACaller initial_msr", initial_msr
)
1394 # "fake program counter" mode (for unit testing)
1398 if isinstance(initial_mem
, tuple):
1399 self
.fake_pc
= initial_mem
[0]
1400 disasm_start
= self
.fake_pc
1402 disasm_start
= initial_pc
1404 # disassembly: we need this for now (not given from the decoder)
1405 self
.disassembly
= {}
1407 for i
, code
in enumerate(disassembly
):
1408 self
.disassembly
[i
*4 + disasm_start
] = code
1410 # set up registers, instruction memory, data memory, PC, SPRs, MSR, CR
1411 self
.svp64rm
= SVP64RM()
1412 if initial_svstate
is None:
1414 if isinstance(initial_svstate
, int):
1415 initial_svstate
= SVP64State(initial_svstate
)
1416 # SVSTATE, MSR and PC
1417 StepLoop
.__init
__(self
, initial_svstate
)
1418 self
.msr
= SelectableInt(initial_msr
, 64) # underlying reg
1420 # GPR FPR SPR registers
1421 initial_sprs
= deepcopy(initial_sprs
) # so as not to get modified
1422 self
.gpr
= GPR(decoder2
, self
, self
.svstate
, regfile
)
1423 self
.fpr
= GPR(decoder2
, self
, self
.svstate
, fpregfile
)
1424 # initialise SPRs before MMU
1425 self
.spr
= SPR(decoder2
, initial_sprs
, gpr
=self
.gpr
)
1427 # set up 4 dummy SVSHAPEs if they aren't already set up
1429 sname
= 'SVSHAPE%d' % i
1430 val
= self
.spr
.get(sname
, 0)
1431 # make sure it's an SVSHAPE -- conversion done by SPR.__setitem__
1432 self
.spr
[sname
] = val
1433 self
.last_op_svshape
= False
1437 self
.mem
= MemMMap(row_bytes
=8,
1438 initial_mem
=initial_mem
,
1440 emulating_mmap
=emulating_mmap
)
1441 self
.imem
= self
.mem
1442 lelf
= self
.mem
.initialize(row_bytes
=4, initial_mem
=initial_insns
)
1443 if isinstance(lelf
, LoadedELF
): # stuff parsed from ELF
1444 initial_pc
= lelf
.pc
1445 for k
, v
in lelf
.gprs
.items():
1446 self
.gpr
[k
] = SelectableInt(v
, 64)
1447 initial_fpscr
= lelf
.fpscr
1448 self
.mem
.log_fancy(kind
=LogType
.InstrInOuts
)
1450 self
.mem
= Mem(row_bytes
=8, initial_mem
=initial_mem
,
1452 self
.mem
.log_fancy(kind
=LogType
.InstrInOuts
)
1453 self
.imem
= Mem(row_bytes
=4, initial_mem
=initial_insns
)
1454 # MMU mode, redirect underlying Mem through RADIX
1456 self
.mem
= RADIX(self
.mem
, self
)
1458 self
.imem
= RADIX(self
.imem
, self
)
1460 # TODO, needed here:
1461 # FPR (same as GPR except for FP nums)
1462 # 4.2.2 p124 FPSCR (definitely "separate" - not in SPR)
1463 # note that mffs, mcrfs, mtfsf "manage" this FPSCR
1464 self
.fpscr
= FPSCRState(initial_fpscr
)
1466 # 2.3.1 CR (and sub-fields CR0..CR6 - CR0 SO comes from XER.SO)
1467 # note that mfocrf, mfcr, mtcr, mtocrf, mcrxrx "manage" CRs
1469 # 2.3.2 LR (actually SPR #8) -- Done
1470 # 2.3.3 CTR (actually SPR #9) -- Done
1471 # 2.3.4 TAR (actually SPR #815)
1472 # 3.2.2 p45 XER (actually SPR #1) -- Done
1473 # 3.2.3 p46 p232 VRSAVE (actually SPR #256)
1475 # create CR then allow portions of it to be "selectable" (below)
1476 self
.cr_fields
= CRFields(initial_cr
)
1477 self
.cr
= self
.cr_fields
.cr
1478 self
.cr_backup
= 0 # sigh, dreadful hack: for fail-first (VLi)
1480 # "undefined", just set to variable-bit-width int (use exts "max")
1481 # self.undefined = SelectableInt(0, EFFECTIVELY_UNLIMITED)
1484 self
.namespace
.update(self
.spr
)
1485 self
.namespace
.update({'GPR': self
.gpr
,
1489 'memassign': self
.memassign
,
1492 'SVSTATE': self
.svstate
,
1493 'SVSHAPE0': self
.spr
['SVSHAPE0'],
1494 'SVSHAPE1': self
.spr
['SVSHAPE1'],
1495 'SVSHAPE2': self
.spr
['SVSHAPE2'],
1496 'SVSHAPE3': self
.spr
['SVSHAPE3'],
1499 'FPSCR': self
.fpscr
,
1500 'undefined': undefined
,
1501 'mode_is_64bit': True,
1502 'SO': XER_bits
['SO'],
1503 'XLEN': 64, # elwidth overrides
1507 if real_page_size
is None:
1508 # PowerISA v3.1B Book III Section 6.7 page 1191 (1217)
1509 # defines real page size as 2 ** 12 bytes (4KiB)
1510 real_page_size
= 2 ** 12
1511 self
.real_page_size
= real_page_size
1512 self
.reserve_addr
= SelectableInt(0, self
.XLEN
)
1513 self
.reserve
= SelectableInt(0, 1)
1514 self
.reserve_length
= SelectableInt(0, 4)
1516 self
.namespace
.update({'RESERVE': self
.RESERVE
,
1517 'RESERVE_ADDR': self
.RESERVE_ADDR
,
1518 'RESERVE_LENGTH': self
.RESERVE_LENGTH
,
1519 'REAL_PAGE_SIZE': self
.REAL_PAGE_SIZE
,
1522 for name
in BFP_FLAG_NAMES
:
1523 setattr(self
, name
, 0)
1525 # update pc to requested start point
1526 self
.set_pc(initial_pc
)
1528 # field-selectable versions of Condition Register
1529 self
.crl
= self
.cr_fields
.crl
1531 self
.namespace
["CR%d" % i
] = self
.crl
[i
]
1533 self
.decoder
= decoder2
.dec
1534 self
.dec2
= decoder2
1536 super().__init
__(XLEN
=self
.namespace
["XLEN"], FPSCR
=self
.fpscr
)
1538 def trace(self
, out
):
1539 if self
.insnlog
is None:
1541 self
.insnlog
.write(out
)
1545 return self
.namespace
["XLEN"]
1552 def RESERVE_LENGTH(self
):
1553 return self
.reserve_length
1556 def RESERVE_ADDR(self
):
1557 return self
.reserve_addr
1560 def REAL_PAGE_SIZE(self
):
1561 return self
.real_page_size
1563 def real_addr(self
, EA
):
1564 """ get the "real address to which `EA` maps"
1566 Specified in PowerISA v3.1B Book II Section 1.7.2.1 page 1049 (1075)
1568 # FIXME: translate EA to a physical address
1575 def call_trap(self
, trap_addr
, trap_bit
):
1576 """calls TRAP and sets up NIA to the new execution location.
1577 next instruction will begin at trap_addr.
1579 self
.TRAP(trap_addr
, trap_bit
)
1580 self
.namespace
['NIA'] = self
.trap_nia
1581 self
.pc
.update(self
.namespace
, self
.is_svp64_mode
)
1583 def TRAP(self
, trap_addr
=0x700, trap_bit
=PIb
.TRAP
):
1584 """TRAP> saves PC, MSR (and TODO SVSTATE), and updates MSR
1586 TRAP function is callable from inside the pseudocode itself,
1587 hence the default arguments. when calling from inside ISACaller
1588 it is best to use call_trap()
1590 trap_addr: int | SelectableInt
1591 the address to go to (before any modifications from `KAIVB`)
1592 trap_bit: int | None
1593 the bit in `SRR1` to set, `None` means don't set any bits.
1595 if isinstance(trap_addr
, SelectableInt
):
1596 trap_addr
= trap_addr
.value
1597 # https://bugs.libre-soc.org/show_bug.cgi?id=859
1598 kaivb
= self
.spr
['KAIVB'].value
1599 msr
= self
.namespace
['MSR'].value
1600 log("TRAP:", hex(trap_addr
), hex(msr
), "kaivb", hex(kaivb
))
1601 # store CIA(+4?) in SRR0, set NIA to 0x700
1602 # store MSR in SRR1, set MSR to um errr something, have to check spec
1603 # store SVSTATE (if enabled) in SVSRR0
1604 self
.spr
['SRR0'].value
= self
.pc
.CIA
.value
1605 self
.spr
['SRR1'].value
= msr
1606 if self
.is_svp64_mode
:
1607 self
.spr
['SVSRR0'] = self
.namespace
['SVSTATE'].value
1608 self
.trap_nia
= SelectableInt(trap_addr |
(kaivb
& ~
0x1fff), 64)
1609 if trap_bit
is not None:
1610 self
.spr
['SRR1'][trap_bit
] = 1 # change *copy* of MSR in SRR1
1612 # set exception bits. TODO: this should, based on the address
1613 # in figure 66 p1065 V3.0B and the table figure 65 p1063 set these
1614 # bits appropriately. however it turns out that *for now* in all
1615 # cases (all trap_addrs) the exact same thing is needed.
1616 self
.msr
[MSRb
.IR
] = 0
1617 self
.msr
[MSRb
.DR
] = 0
1618 self
.msr
[MSRb
.FE0
] = 0
1619 self
.msr
[MSRb
.FE1
] = 0
1620 self
.msr
[MSRb
.EE
] = 0
1621 self
.msr
[MSRb
.RI
] = 0
1622 self
.msr
[MSRb
.SF
] = 1
1623 self
.msr
[MSRb
.TM
] = 0
1624 self
.msr
[MSRb
.VEC
] = 0
1625 self
.msr
[MSRb
.VSX
] = 0
1626 self
.msr
[MSRb
.PR
] = 0
1627 self
.msr
[MSRb
.FP
] = 0
1628 self
.msr
[MSRb
.PMM
] = 0
1629 self
.msr
[MSRb
.TEs
] = 0
1630 self
.msr
[MSRb
.TEe
] = 0
1631 self
.msr
[MSRb
.UND
] = 0
1632 self
.msr
[MSRb
.LE
] = 1
1634 def memassign(self
, ea
, sz
, val
):
1635 self
.mem
.memassign(ea
, sz
, val
)
1637 def prep_namespace(self
, insn_name
, info
, xlen
):
1638 # TODO: get field names from form in decoder*1* (not decoder2)
1639 # decoder2 is hand-created, and decoder1.sigform is auto-generated
1641 # then "yield" fields only from op_fields rather than hard-coded
1643 formname
, op_fields
= info
.form
, info
.op_fields
1644 fields
= self
.decoder
.sigforms
[formname
]
1645 log("prep_namespace", formname
, op_fields
, insn_name
)
1646 for name
in op_fields
:
1647 # CR immediates. deal with separately. needs modifying
1649 crlen5
= ['BC', 'BA', 'BB', 'BT', 'BI'] # 5-bit
1650 crlen3
= ['BF', 'BFA'] # 3-bit (BF: bit-field)
1651 if self
.is_svp64_mode
and name
in crlen5
:
1652 # 5-bit, must reconstruct the value
1654 regnum
, is_vec
= yield from get_cr_out(self
.dec2
, name
)
1656 regnum
, is_vec
= yield from get_cr_in(self
.dec2
, name
)
1657 sig
= getattr(fields
, name
)
1659 # low 2 LSBs (CR field selector) remain same, CR num extended
1660 assert regnum
<= 7, "sigh, TODO, 128 CR fields"
1661 val
= (val
& 0b11) |
(regnum
<< 2)
1662 elif self
.is_svp64_mode
and name
in crlen3
:
1664 regnum
, is_vec
= yield from get_cr_out(self
.dec2
, name
)
1666 regnum
, is_vec
= yield from get_cr_in(self
.dec2
, name
)
1667 log('hack %s' % name
, regnum
, is_vec
)
1670 sig
= getattr(fields
, name
)
1672 # these are all opcode fields involved in index-selection of CR,
1673 # and need to do "standard" arithmetic. CR[BA+32] for example
1674 # would, if using SelectableInt, only be 5-bit.
1675 if name
not in crlen3
and name
not in crlen5
:
1676 val
= SelectableInt(val
, sig
.width
)
1678 # finally put the field into the namespace
1679 self
.namespace
[name
] = val
1681 self
.namespace
['XER'] = self
.spr
['XER']
1682 self
.namespace
['CA'] = self
.spr
['XER'][XER_bits
['CA']].value
1683 self
.namespace
['CA32'] = self
.spr
['XER'][XER_bits
['CA32']].value
1684 self
.namespace
['OV'] = self
.spr
['XER'][XER_bits
['OV']].value
1685 self
.namespace
['OV32'] = self
.spr
['XER'][XER_bits
['OV32']].value
1686 self
.namespace
['XLEN'] = xlen
1687 self
.namespace
['RESERVE'] = self
.reserve
1688 self
.namespace
['RESERVE_ADDR'] = self
.reserve_addr
1689 self
.namespace
['RESERVE_LENGTH'] = self
.reserve_length
1691 # add some SVSTATE convenience variables
1692 vl
= self
.svstate
.vl
1693 srcstep
= self
.svstate
.srcstep
1694 self
.namespace
['VL'] = vl
1695 self
.namespace
['srcstep'] = srcstep
1697 # take a copy of the CR field value: if non-VLi fail-first fails
1698 # this is because the pseudocode writes *directly* to CR. sigh
1699 self
.cr_backup
= self
.cr
.value
1701 # sv.bc* need some extra fields
1702 if not self
.is_svp64_mode
or not insn_name
.startswith("sv.bc"):
1705 # blegh grab bits manually
1706 mode
= yield self
.dec2
.rm_dec
.rm_in
.mode
1707 # convert to SelectableInt before test
1708 mode
= SelectableInt(mode
, 5)
1709 bc_vlset
= mode
[SVP64MODEb
.BC_VLSET
] != 0
1710 bc_vli
= mode
[SVP64MODEb
.BC_VLI
] != 0
1711 bc_snz
= mode
[SVP64MODEb
.BC_SNZ
] != 0
1712 bc_vsb
= yield self
.dec2
.rm_dec
.bc_vsb
1713 bc_ctrtest
= yield self
.dec2
.rm_dec
.bc_ctrtest
1714 bc_lru
= yield self
.dec2
.rm_dec
.bc_lru
1715 bc_gate
= yield self
.dec2
.rm_dec
.bc_gate
1716 sz
= yield self
.dec2
.rm_dec
.pred_sz
1717 self
.namespace
['mode'] = SelectableInt(mode
, 5)
1718 self
.namespace
['ALL'] = SelectableInt(bc_gate
, 1)
1719 self
.namespace
['VSb'] = SelectableInt(bc_vsb
, 1)
1720 self
.namespace
['LRu'] = SelectableInt(bc_lru
, 1)
1721 self
.namespace
['CTRtest'] = SelectableInt(bc_ctrtest
, 1)
1722 self
.namespace
['VLSET'] = SelectableInt(bc_vlset
, 1)
1723 self
.namespace
['VLI'] = SelectableInt(bc_vli
, 1)
1724 self
.namespace
['sz'] = SelectableInt(sz
, 1)
1725 self
.namespace
['SNZ'] = SelectableInt(bc_snz
, 1)
1727 def get_kludged_op_add_ca_ov(self
, inputs
, inp_ca_ov
):
1728 """ this was not at all necessary to do. this function massively
1729 duplicates - in a laborious and complex fashion - the contents of
1730 the CSV files that were extracted two years ago from microwatt's
1731 source code. A-inversion is the "inv A" column, output inversion
1732 is the "inv out" column, carry-in equal to 0 or 1 or CA is the
1735 all of that information is available in
1736 self.instrs[ins_name].op_fields
1737 where info is usually assigned to self.instrs[ins_name]
1739 https://git.libre-soc.org/?p=openpower-isa.git;a=blob;f=openpower/isatables/minor_31.csv;hb=HEAD
1741 the immediate constants are *also* decoded correctly and placed
1742 usually by DecodeIn2Imm into operand2, as part of power_decoder2.py
1744 def ca(a
, b
, ca_in
, width
):
1745 mask
= (1 << width
) - 1
1746 y
= (a
& mask
) + (b
& mask
) + ca_in
1749 asmcode
= yield self
.dec2
.dec
.op
.asmcode
1750 insn
= insns
.get(asmcode
)
1751 SI
= yield self
.dec2
.dec
.SI
1754 inputs
= [i
.value
for i
in inputs
]
1757 if insn
in ("add", "addo", "addc", "addco"):
1761 elif insn
== "addic" or insn
== "addic.":
1765 elif insn
in ("subf", "subfo", "subfc", "subfco"):
1769 elif insn
== "subfic":
1773 elif insn
== "adde" or insn
== "addeo":
1777 elif insn
== "subfe" or insn
== "subfeo":
1781 elif insn
== "addme" or insn
== "addmeo":
1785 elif insn
== "addze" or insn
== "addzeo":
1789 elif insn
== "subfme" or insn
== "subfmeo":
1793 elif insn
== "subfze" or insn
== "subfzeo":
1797 elif insn
== "addex":
1798 # CA[32] aren't actually written, just generate so we have
1799 # something to return
1800 ca64
= ov64
= ca(inputs
[0], inputs
[1], OV
, 64)
1801 ca32
= ov32
= ca(inputs
[0], inputs
[1], OV
, 32)
1802 return ca64
, ca32
, ov64
, ov32
1803 elif insn
== "neg" or insn
== "nego":
1808 raise NotImplementedError(
1809 "op_add kludge unimplemented instruction: ", asmcode
, insn
)
1811 ca64
= ca(a
, b
, ca_in
, 64)
1812 ca32
= ca(a
, b
, ca_in
, 32)
1813 ov64
= ca64
!= ca(a
, b
, ca_in
, 63)
1814 ov32
= ca32
!= ca(a
, b
, ca_in
, 31)
1815 return ca64
, ca32
, ov64
, ov32
1817 def handle_carry_(self
, inputs
, output
, ca
, ca32
, inp_ca_ov
):
1818 if ca
is not None and ca32
is not None:
1820 op
= yield self
.dec2
.e
.do
.insn_type
1821 if op
== MicrOp
.OP_ADD
.value
and ca
is None and ca32
is None:
1822 retval
= yield from self
.get_kludged_op_add_ca_ov(
1824 ca
, ca32
, ov
, ov32
= retval
1825 asmcode
= yield self
.dec2
.dec
.op
.asmcode
1826 if insns
.get(asmcode
) == 'addex':
1827 # TODO: if 32-bit mode, set ov to ov32
1828 self
.spr
['XER'][XER_bits
['OV']] = ov
1829 self
.spr
['XER'][XER_bits
['OV32']] = ov32
1830 log(f
"write OV/OV32 OV={ov} OV32={ov32}",
1831 kind
=LogType
.InstrInOuts
)
1833 # TODO: if 32-bit mode, set ca to ca32
1834 self
.spr
['XER'][XER_bits
['CA']] = ca
1835 self
.spr
['XER'][XER_bits
['CA32']] = ca32
1836 log(f
"write CA/CA32 CA={ca} CA32={ca32}",
1837 kind
=LogType
.InstrInOuts
)
1839 inv_a
= yield self
.dec2
.e
.do
.invert_in
1841 inputs
[0] = ~inputs
[0]
1843 imm_ok
= yield self
.dec2
.e
.do
.imm_data
.ok
1845 imm
= yield self
.dec2
.e
.do
.imm_data
.data
1846 inputs
.append(SelectableInt(imm
, 64))
1849 log("gt input", x
, output
)
1850 gt
= (gtu(x
, output
))
1853 cy
= 1 if any(gts
) else 0
1855 if ca
is None: # already written
1856 self
.spr
['XER'][XER_bits
['CA']] = cy
1859 # ARGH... different for OP_ADD... *sigh*...
1860 op
= yield self
.dec2
.e
.do
.insn_type
1861 if op
== MicrOp
.OP_ADD
.value
:
1862 res32
= (output
.value
& (1 << 32)) != 0
1863 a32
= (inputs
[0].value
& (1 << 32)) != 0
1864 if len(inputs
) >= 2:
1865 b32
= (inputs
[1].value
& (1 << 32)) != 0
1868 cy32
= res32 ^ a32 ^ b32
1869 log("CA32 ADD", cy32
)
1873 log("input", x
, output
)
1874 log(" x[32:64]", x
, x
[32:64])
1875 log(" o[32:64]", output
, output
[32:64])
1876 gt
= (gtu(x
[32:64], output
[32:64])) == SelectableInt(1, 1)
1878 cy32
= 1 if any(gts
) else 0
1879 log("CA32", cy32
, gts
)
1880 if ca32
is None: # already written
1881 self
.spr
['XER'][XER_bits
['CA32']] = cy32
1883 def handle_overflow(self
, inputs
, output
, div_overflow
, inp_ca_ov
):
1884 op
= yield self
.dec2
.e
.do
.insn_type
1885 if op
== MicrOp
.OP_ADD
.value
:
1886 retval
= yield from self
.get_kludged_op_add_ca_ov(
1888 ca
, ca32
, ov
, ov32
= retval
1889 # TODO: if 32-bit mode, set ov to ov32
1890 self
.spr
['XER'][XER_bits
['OV']] = ov
1891 self
.spr
['XER'][XER_bits
['OV32']] = ov32
1892 self
.spr
['XER'][XER_bits
['SO']] |
= ov
1894 if hasattr(self
.dec2
.e
.do
, "invert_in"):
1895 inv_a
= yield self
.dec2
.e
.do
.invert_in
1897 inputs
[0] = ~inputs
[0]
1899 imm_ok
= yield self
.dec2
.e
.do
.imm_data
.ok
1901 imm
= yield self
.dec2
.e
.do
.imm_data
.data
1902 inputs
.append(SelectableInt(imm
, 64))
1903 log("handle_overflow", inputs
, output
, div_overflow
)
1904 if len(inputs
) < 2 and div_overflow
is None:
1907 # div overflow is different: it's returned by the pseudo-code
1908 # because it's more complex than can be done by analysing the output
1909 if div_overflow
is not None:
1910 ov
, ov32
= div_overflow
, div_overflow
1911 # arithmetic overflow can be done by analysing the input and output
1912 elif len(inputs
) >= 2:
1914 input_sgn
= [exts(x
.value
, x
.bits
) < 0 for x
in inputs
]
1915 output_sgn
= exts(output
.value
, output
.bits
) < 0
1916 ov
= 1 if input_sgn
[0] == input_sgn
[1] and \
1917 output_sgn
!= input_sgn
[0] else 0
1920 input32_sgn
= [exts(x
.value
, 32) < 0 for x
in inputs
]
1921 output32_sgn
= exts(output
.value
, 32) < 0
1922 ov32
= 1 if input32_sgn
[0] == input32_sgn
[1] and \
1923 output32_sgn
!= input32_sgn
[0] else 0
1925 # now update XER OV/OV32/SO
1926 so
= self
.spr
['XER'][XER_bits
['SO']]
1927 new_so
= so | ov
# sticky overflow ORs in old with new
1928 self
.spr
['XER'][XER_bits
['OV']] = ov
1929 self
.spr
['XER'][XER_bits
['OV32']] = ov32
1930 self
.spr
['XER'][XER_bits
['SO']] = new_so
1931 log(" set overflow", ov
, ov32
, so
, new_so
)
1933 def handle_comparison(self
, out
, cr_idx
=0, overflow
=None, no_so
=False):
1934 assert isinstance(out
, SelectableInt
), \
1935 "out zero not a SelectableInt %s" % repr(outputs
)
1936 log("handle_comparison", out
.bits
, hex(out
.value
))
1937 # TODO - XXX *processor* in 32-bit mode
1938 # https://bugs.libre-soc.org/show_bug.cgi?id=424
1940 # o32 = exts(out.value, 32)
1941 # print ("handle_comparison exts 32 bit", hex(o32))
1942 out
= exts(out
.value
, out
.bits
)
1943 log("handle_comparison exts", hex(out
))
1944 # create the three main CR flags, EQ GT LT
1945 zero
= SelectableInt(out
== 0, 1)
1946 positive
= SelectableInt(out
> 0, 1)
1947 negative
= SelectableInt(out
< 0, 1)
1948 # get (or not) XER.SO. for setvl this is important *not* to read SO
1950 SO
= SelectableInt(1, 0)
1952 SO
= self
.spr
['XER'][XER_bits
['SO']]
1953 log("handle_comparison SO", SO
.value
,
1954 "overflow", overflow
,
1956 "+ve", positive
.value
,
1957 "-ve", negative
.value
)
1958 # alternative overflow checking (setvl mainly at the moment)
1959 if overflow
is not None and overflow
== 1:
1960 SO
= SelectableInt(1, 1)
1961 # create the four CR field values and set the required CR field
1962 cr_field
= selectconcat(negative
, positive
, zero
, SO
)
1963 log("handle_comparison cr_field", self
.cr
, cr_idx
, cr_field
)
1964 self
.crl
[cr_idx
].eq(cr_field
)
1967 def set_pc(self
, pc_val
):
1968 self
.namespace
['NIA'] = SelectableInt(pc_val
, 64)
1969 self
.pc
.update(self
.namespace
, self
.is_svp64_mode
)
1971 def get_next_insn(self
):
1972 """check instruction
1975 pc
= self
.pc
.CIA
.value
1978 ins
= self
.imem
.ld(pc
, 4, False, True, instr_fetch
=True)
1980 raise KeyError("no instruction at 0x%x" % pc
)
1983 def setup_one(self
):
1984 """set up one instruction
1986 pc
, insn
= self
.get_next_insn()
1987 yield from self
.setup_next_insn(pc
, insn
)
1989 # cache since it's really slow to construct
1990 __PREFIX_CACHE
= SVP64Instruction
.Prefix(SelectableInt(value
=0, bits
=32))
1992 def __decode_prefix(self
, opcode
):
1993 pfx
= self
.__PREFIX
_CACHE
1994 pfx
.storage
.eq(opcode
)
1997 def setup_next_insn(self
, pc
, ins
):
1998 """set up next instruction
2001 log("setup: 0x%x 0x%x %s" % (pc
, ins
& 0xffffffff, bin(ins
)))
2002 log("CIA NIA", self
.respect_pc
, self
.pc
.CIA
.value
, self
.pc
.NIA
.value
)
2004 yield self
.dec2
.sv_rm
.eq(0)
2005 yield self
.dec2
.dec
.raw_opcode_in
.eq(ins
& 0xffffffff)
2006 yield self
.dec2
.dec
.bigendian
.eq(self
.bigendian
)
2007 yield self
.dec2
.state
.msr
.eq(self
.msr
.value
)
2008 yield self
.dec2
.state
.pc
.eq(pc
)
2009 if self
.svstate
is not None:
2010 yield self
.dec2
.state
.svstate
.eq(self
.svstate
.value
)
2012 # SVP64. first, check if the opcode is EXT001, and SVP64 id bits set
2014 opcode
= yield self
.dec2
.dec
.opcode_in
2015 opcode
= SelectableInt(value
=opcode
, bits
=32)
2016 pfx
= self
.__decode
_prefix
(opcode
)
2017 log("prefix test: opcode:", pfx
.PO
, bin(pfx
.PO
), pfx
.id)
2018 self
.is_svp64_mode
= bool((pfx
.PO
== 0b000001) and (pfx
.id == 0b11))
2019 self
.pc
.update_nia(self
.is_svp64_mode
)
2021 yield self
.dec2
.is_svp64_mode
.eq(self
.is_svp64_mode
)
2022 self
.namespace
['NIA'] = self
.pc
.NIA
2023 self
.namespace
['SVSTATE'] = self
.svstate
2024 if not self
.is_svp64_mode
:
2027 # in SVP64 mode. decode/print out svp64 prefix, get v3.0B instruction
2028 log("svp64.rm", bin(pfx
.rm
))
2029 log(" svstate.vl", self
.svstate
.vl
)
2030 log(" svstate.mvl", self
.svstate
.maxvl
)
2031 ins
= self
.imem
.ld(pc
+4, 4, False, True, instr_fetch
=True)
2032 log(" svsetup: 0x%x 0x%x %s" % (pc
+4, ins
& 0xffffffff, bin(ins
)))
2033 yield self
.dec2
.dec
.raw_opcode_in
.eq(ins
& 0xffffffff) # v3.0B suffix
2034 yield self
.dec2
.sv_rm
.eq(int(pfx
.rm
)) # svp64 prefix
2037 def execute_one(self
):
2038 """execute one instruction
2040 # get the disassembly code for this instruction
2041 if not self
.disassembly
:
2042 code
= yield from self
.get_assembly_name()
2045 if self
.is_svp64_mode
:
2046 offs
, dbg
= 4, "svp64 "
2047 code
= self
.disassembly
[self
._pc
+offs
]
2048 log(" %s sim-execute" % dbg
, hex(self
._pc
), code
)
2049 opname
= code
.split(' ')[0]
2051 asmop
= yield from self
.call(opname
) # execute the instruction
2052 except MemException
as e
: # check for memory errors
2053 if e
.args
[0] == 'unaligned': # alignment error
2054 # run a Trap but set DAR first
2055 print("memory unaligned exception, DAR", e
.dar
, repr(e
))
2056 self
.spr
['DAR'] = SelectableInt(e
.dar
, 64)
2057 self
.call_trap(0x600, PIb
.PRIV
) # 0x600, privileged
2059 elif e
.args
[0] == 'invalid': # invalid
2060 # run a Trap but set DAR first
2061 log("RADIX MMU memory invalid error, mode %s" % e
.mode
)
2062 if e
.mode
== 'EXECUTE':
2063 # XXX TODO: must set a few bits in SRR1,
2064 # see microwatt loadstore1.vhdl
2065 # if m_in.segerr = '0' then
2066 # v.srr1(47 - 33) := m_in.invalid;
2067 # v.srr1(47 - 35) := m_in.perm_error; -- noexec fault
2068 # v.srr1(47 - 44) := m_in.badtree;
2069 # v.srr1(47 - 45) := m_in.rc_error;
2070 # v.intr_vec := 16#400#;
2072 # v.intr_vec := 16#480#;
2073 self
.call_trap(0x400, PIb
.PRIV
) # 0x400, privileged
2075 self
.call_trap(0x300, PIb
.PRIV
) # 0x300, privileged
2077 # not supported yet:
2078 raise e
# ... re-raise
2080 # append to the trace log file
2081 self
.trace(" # %s %s\n" % (asmop
, code
))
2083 log("gprs after insn %s - code" % asmop
, code
)
2086 for i
in range(len(self
.crl
)):
2087 crs
.append(bin(self
.crl
[i
].asint()))
2088 log("crs", " ".join(crs
))
2089 log("vl,maxvl", self
.svstate
.vl
, self
.svstate
.maxvl
)
2091 # don't use this except in special circumstances
2092 if not self
.respect_pc
:
2095 log("execute one, CIA NIA", hex(self
.pc
.CIA
.value
),
2096 hex(self
.pc
.NIA
.value
))
2098 def get_assembly_name(self
):
2099 # TODO, asmregs is from the spec, e.g. add RT,RA,RB
2100 # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
2101 dec_insn
= yield self
.dec2
.e
.do
.insn
2102 insn_1_11
= yield self
.dec2
.e
.do
.insn
[1:11]
2103 asmcode
= yield self
.dec2
.dec
.op
.asmcode
2104 int_op
= yield self
.dec2
.dec
.op
.internal_op
2105 log("get assembly name asmcode", asmcode
, int_op
,
2106 hex(dec_insn
), bin(insn_1_11
))
2107 asmop
= insns
.get(asmcode
, None)
2109 # sigh reconstruct the assembly instruction name
2110 if hasattr(self
.dec2
.e
.do
, "oe"):
2111 ov_en
= yield self
.dec2
.e
.do
.oe
.oe
2112 ov_ok
= yield self
.dec2
.e
.do
.oe
.ok
2116 if hasattr(self
.dec2
.e
.do
, "rc"):
2117 rc_en
= yield self
.dec2
.e
.do
.rc
.rc
2118 rc_ok
= yield self
.dec2
.e
.do
.rc
.ok
2122 # annoying: ignore rc_ok if RC1 is set (for creating *assembly name*)
2123 RC1
= yield self
.dec2
.rm_dec
.RC1
2127 # grrrr have to special-case MUL op (see DecodeOE)
2128 log("ov %d en %d rc %d en %d op %d" %
2129 (ov_ok
, ov_en
, rc_ok
, rc_en
, int_op
))
2130 if int_op
in [MicrOp
.OP_MUL_H64
.value
, MicrOp
.OP_MUL_H32
.value
]:
2135 if not asmop
.endswith("."): # don't add "." to "andis."
2138 if hasattr(self
.dec2
.e
.do
, "lk"):
2139 lk
= yield self
.dec2
.e
.do
.lk
2142 log("int_op", int_op
)
2143 if int_op
in [MicrOp
.OP_B
.value
, MicrOp
.OP_BC
.value
]:
2144 AA
= yield self
.dec2
.dec
.fields
.FormI
.AA
[0:-1]
2148 spr_msb
= yield from self
.get_spr_msb()
2149 if int_op
== MicrOp
.OP_MFCR
.value
:
2154 # XXX TODO: for whatever weird reason this doesn't work
2155 # https://bugs.libre-soc.org/show_bug.cgi?id=390
2156 if int_op
== MicrOp
.OP_MTCRF
.value
:
2163 def reset_remaps(self
):
2164 self
.remap_loopends
= [0] * 4
2165 self
.remap_idxs
= [0, 1, 2, 3]
2167 def get_remap_indices(self
):
2168 """WARNING, this function stores remap_idxs and remap_loopends
2169 in the class for later use. this to avoid problems with yield
2171 # go through all iterators in lock-step, advance to next remap_idx
2172 srcstep
, dststep
, ssubstep
, dsubstep
= self
.get_src_dststeps()
2173 # get four SVSHAPEs. here we are hard-coding
2175 SVSHAPE0
= self
.spr
['SVSHAPE0']
2176 SVSHAPE1
= self
.spr
['SVSHAPE1']
2177 SVSHAPE2
= self
.spr
['SVSHAPE2']
2178 SVSHAPE3
= self
.spr
['SVSHAPE3']
2179 # set up the iterators
2180 remaps
= [(SVSHAPE0
, SVSHAPE0
.get_iterator()),
2181 (SVSHAPE1
, SVSHAPE1
.get_iterator()),
2182 (SVSHAPE2
, SVSHAPE2
.get_iterator()),
2183 (SVSHAPE3
, SVSHAPE3
.get_iterator()),
2187 for i
, (shape
, remap
) in enumerate(remaps
):
2188 # zero is "disabled"
2189 if shape
.value
== 0x0:
2190 self
.remap_idxs
[i
] = 0
2191 # pick src or dststep depending on reg num (0-2=in, 3-4=out)
2192 step
= dststep
if (i
in [3, 4]) else srcstep
2193 # this is terrible. O(N^2) looking for the match. but hey.
2194 for idx
, (remap_idx
, loopends
) in enumerate(remap
):
2197 self
.remap_idxs
[i
] = remap_idx
2198 self
.remap_loopends
[i
] = loopends
2199 dbg
.append((i
, step
, remap_idx
, loopends
))
2200 for (i
, step
, remap_idx
, loopends
) in dbg
:
2201 log("SVSHAPE %d idx, end" % i
, step
, remap_idx
, bin(loopends
))
2204 def get_spr_msb(self
):
2205 dec_insn
= yield self
.dec2
.e
.do
.insn
2206 return dec_insn
& (1 << 20) != 0 # sigh - XFF.spr[-1]?
2208 def call(self
, name
, syscall_emu_active
=False):
2209 """call(opcode) - the primary execution point for instructions
2211 self
.last_st_addr
= None # reset the last known store address
2212 self
.last_ld_addr
= None # etc.
2214 ins_name
= name
.strip() # remove spaces if not already done so
2216 log("halted - not executing", ins_name
)
2219 # TODO, asmregs is from the spec, e.g. add RT,RA,RB
2220 # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
2221 asmop
= yield from self
.get_assembly_name()
2222 log("call", ins_name
, asmop
, kind
=LogType
.InstrInOuts
)
2224 # sv.setvl is *not* a loop-function. sigh
2225 log("is_svp64_mode", self
.is_svp64_mode
, asmop
)
2228 int_op
= yield self
.dec2
.dec
.op
.internal_op
2229 spr_msb
= yield from self
.get_spr_msb()
2231 instr_is_privileged
= False
2232 if int_op
in [MicrOp
.OP_ATTN
.value
,
2233 MicrOp
.OP_MFMSR
.value
,
2234 MicrOp
.OP_MTMSR
.value
,
2235 MicrOp
.OP_MTMSRD
.value
,
2237 MicrOp
.OP_RFID
.value
]:
2238 instr_is_privileged
= True
2239 if int_op
in [MicrOp
.OP_MFSPR
.value
,
2240 MicrOp
.OP_MTSPR
.value
] and spr_msb
:
2241 instr_is_privileged
= True
2243 # check MSR priv bit and whether op is privileged: if so, throw trap
2244 PR
= self
.msr
[MSRb
.PR
]
2245 log("is priv", instr_is_privileged
, hex(self
.msr
.value
), PR
)
2246 if instr_is_privileged
and PR
== 1:
2247 self
.call_trap(0x700, PIb
.PRIV
)
2250 # check halted condition
2251 if ins_name
== 'attn':
2255 # User mode system call emulation consists of several steps:
2256 # 1. Detect whether instruction is sc or scv.
2257 # 2. Call the HDL implementation which invokes trap.
2258 # 3. Reroute the guest system call to host system call.
2259 # 4. Force return from the interrupt as if we had guest OS.
2260 # FIXME: enable PPC_FEATURE2_SCV in mem.py DEFAULT_AT_HWCAP2 when
2261 # scv emulation works.
2262 if ((asmop
in ("sc", "scv")) and
2263 (self
.syscall
is not None) and
2264 not syscall_emu_active
):
2265 # Memoize PC and trigger an interrupt
2267 pc
= self
.pc
.CIA
.value
2270 yield from self
.call(asmop
, syscall_emu_active
=True)
2272 # Reroute the syscall to host OS
2273 identifier
= self
.gpr(0)
2274 arguments
= map(self
.gpr
, range(3, 9))
2275 result
= self
.syscall(identifier
, *arguments
)
2276 self
.gpr
.write(3, result
, False, self
.namespace
["XLEN"])
2278 # Return from interrupt
2279 yield from self
.call("rfid", syscall_emu_active
=True)
2281 elif ((name
in ("rfid", "hrfid")) and syscall_emu_active
):
2284 # check illegal instruction
2286 if ins_name
not in ['mtcrf', 'mtocrf']:
2287 illegal
= ins_name
!= asmop
2289 # list of instructions not being supported by binutils (.long)
2290 dotstrp
= asmop
[:-1] if asmop
[-1] == '.' else asmop
2291 if dotstrp
in [*FPTRANS_INSNS
,
2293 'ffmadds', 'fdmadds', 'ffadds',
2295 "brh", "brw", "brd",
2296 'setvl', 'svindex', 'svremap', 'svstep',
2297 'svshape', 'svshape2',
2298 'binlog', 'crbinlog', 'crfbinlog',
2299 'crternlogi', 'crfternlogi', 'ternlogi',
2300 'bmask', 'cprop', 'gbbd',
2301 'absdu', 'absds', 'absdacs', 'absdacu', 'avgadd',
2302 'fmvis', 'fishmv', 'pcdec', "maddedu", "divmod2du",
2303 "dsld", "dsrd", "maddedus",
2304 "sadd", "saddw", "sadduw",
2309 "maddsubrs", "maddrs", "msubrs",
2310 "cfuged", "cntlzdm", "cnttzdm", "pdepd", "pextd",
2311 "setbc", "setbcr", "setnbc", "setnbcr",
2316 # match against instructions treated as nop, see nop below
2317 if asmop
.startswith("dcbt"):
2321 # branch-conditional redirects to sv.bc
2322 if asmop
.startswith('bc') and self
.is_svp64_mode
:
2323 ins_name
= 'sv.%s' % ins_name
2325 # ld-immediate-with-pi mode redirects to ld-with-postinc
2326 ldst_imm_postinc
= False
2327 if 'u' in ins_name
and self
.is_svp64_mode
:
2328 ldst_pi
= yield self
.dec2
.rm_dec
.ldst_postinc
2330 ins_name
= ins_name
.replace("u", "up")
2331 ldst_imm_postinc
= True
2332 log(" enable ld/st postinc", ins_name
)
2334 log(" post-processed name", dotstrp
, ins_name
, asmop
)
2336 # illegal instructions call TRAP at 0x700
2338 print("illegal", ins_name
, asmop
)
2339 self
.call_trap(0x700, PIb
.ILLEG
)
2340 print("name %s != %s - calling ILLEGAL trap, PC: %x" %
2341 (ins_name
, asmop
, self
.pc
.CIA
.value
))
2344 # this is for setvl "Vertical" mode: if set true,
2345 # srcstep/dststep is explicitly advanced. mode says which SVSTATE to
2346 # test for Rc=1 end condition. 3 bits of all 3 loops are put into CR0
2347 self
.allow_next_step_inc
= False
2348 self
.svstate_next_mode
= 0
2350 # nop has to be supported, we could let the actual op calculate
2351 # but PowerDecoder has a pattern for nop
2352 if ins_name
== 'nop':
2353 self
.update_pc_next()
2356 # get elwidths, defaults to 64
2360 if self
.is_svp64_mode
:
2361 ew_src
= yield self
.dec2
.rm_dec
.ew_src
2362 ew_dst
= yield self
.dec2
.rm_dec
.ew_dst
2363 ew_src
= 8 << (3-int(ew_src
)) # convert to bitlength
2364 ew_dst
= 8 << (3-int(ew_dst
)) # convert to bitlength
2365 xlen
= max(ew_src
, ew_dst
)
2366 log("elwidth", ew_src
, ew_dst
)
2367 log("XLEN:", self
.is_svp64_mode
, xlen
)
2369 # look up instruction in ISA.instrs, prepare namespace
2370 if ins_name
== 'pcdec': # grrrr yes there are others ("stbcx." etc.)
2371 info
= self
.instrs
[ins_name
+"."]
2372 elif asmop
[-1] == '.' and asmop
in self
.instrs
:
2373 info
= self
.instrs
[asmop
]
2375 info
= self
.instrs
[ins_name
]
2376 yield from self
.prep_namespace(ins_name
, info
, xlen
)
2378 # dict retains order
2379 inputs
= dict.fromkeys(create_full_args(
2380 read_regs
=info
.read_regs
, special_regs
=info
.special_regs
,
2381 uninit_regs
=info
.uninit_regs
, write_regs
=info
.write_regs
))
2383 # preserve order of register names
2384 write_without_special_regs
= OrderedSet(info
.write_regs
)
2385 write_without_special_regs
-= OrderedSet(info
.special_regs
)
2386 input_names
= create_args([
2387 *info
.read_regs
, *info
.uninit_regs
, *write_without_special_regs
])
2388 log("input names", input_names
)
2390 # get SVP64 entry for the current instruction
2391 sv_rm
= self
.svp64rm
.instrs
.get(ins_name
)
2392 if sv_rm
is not None:
2393 dest_cr
, src_cr
, src_byname
, dest_byname
= decode_extra(sv_rm
)
2395 dest_cr
, src_cr
, src_byname
, dest_byname
= False, False, {}, {}
2396 log("sv rm", sv_rm
, dest_cr
, src_cr
, src_byname
, dest_byname
)
2398 # see if srcstep/dststep need skipping over masked-out predicate bits
2399 # svstep also needs advancement because it calls SVSTATE_NEXT.
2400 # bit the remaps get computed just after pre_inc moves them on
2401 # with remap_set_steps substituting for PowerDecider2 not doing it,
2402 # and SVSTATE_NEXT not being able to.use yield, the preinc on
2403 # svstep is necessary for now.
2405 if (self
.is_svp64_mode
or ins_name
in ['svstep']):
2406 yield from self
.svstate_pre_inc()
2407 if self
.is_svp64_mode
:
2408 pre
= yield from self
.update_new_svstate_steps()
2410 self
.svp64_reset_loop()
2412 self
.update_pc_next()
2414 srcstep
, dststep
, ssubstep
, dsubstep
= self
.get_src_dststeps()
2415 pred_dst_zero
= self
.pred_dst_zero
2416 pred_src_zero
= self
.pred_src_zero
2417 vl
= self
.svstate
.vl
2418 subvl
= yield self
.dec2
.rm_dec
.rm_in
.subvl
2420 # VL=0 in SVP64 mode means "do nothing: skip instruction"
2421 if self
.is_svp64_mode
and vl
== 0:
2422 self
.pc
.update(self
.namespace
, self
.is_svp64_mode
)
2423 log("SVP64: VL=0, end of call", self
.namespace
['CIA'],
2424 self
.namespace
['NIA'], kind
=LogType
.InstrInOuts
)
2427 # for when SVREMAP is active, using pre-arranged schedule.
2428 # note: modifying PowerDecoder2 needs to "settle"
2429 remap_en
= self
.svstate
.SVme
2430 persist
= self
.svstate
.RMpst
2431 active
= (persist
or self
.last_op_svshape
) and remap_en
!= 0
2432 if self
.is_svp64_mode
:
2433 yield self
.dec2
.remap_active
.eq(remap_en
if active
else 0)
2435 if persist
or self
.last_op_svshape
:
2436 remaps
= self
.get_remap_indices()
2437 if self
.is_svp64_mode
and (persist
or self
.last_op_svshape
):
2438 yield from self
.remap_set_steps(remaps
)
2439 # after that, settle down (combinatorial) to let Vector reg numbers
2440 # work themselves out
2442 if self
.is_svp64_mode
:
2443 remap_active
= yield self
.dec2
.remap_active
2445 remap_active
= False
2446 log("remap active", bin(remap_active
), self
.is_svp64_mode
)
2448 # LDST does *not* allow elwidth overrides on RA (Effective Address).
2449 # this has to be detected. XXX TODO: RB for ldst-idx *may* need
2450 # conversion (to 64-bit) also.
2451 # see write reg this *HAS* to also override XLEN to 64 on LDST/Update
2452 sv_mode
= yield self
.dec2
.rm_dec
.sv_mode
2453 is_ldst
= (sv_mode
in [SVMode
.LDST_IDX
.value
, SVMode
.LDST_IMM
.value
] \
2454 and self
.is_svp64_mode
)
2455 log("is_ldst", sv_mode
, is_ldst
)
2457 # main input registers (RT, RA ...)
2458 for name
in input_names
:
2459 if name
== "overflow":
2460 inputs
[name
] = SelectableInt(0, 1)
2461 elif name
.startswith("RESERVE"):
2462 inputs
[name
] = getattr(self
, name
)
2463 elif name
== "FPSCR":
2464 inputs
[name
] = self
.FPSCR
2465 elif name
in ("CA", "CA32", "OV", "OV32"):
2466 inputs
[name
] = self
.spr
['XER'][XER_bits
[name
]]
2468 inputs
[name
] = self
.crl
[0]
2469 elif name
in spr_byname
:
2470 inputs
[name
] = self
.spr
[name
]
2471 elif is_ldst
and name
== 'RA':
2472 regval
= (yield from self
.get_input(name
, ew_src
, 64))
2473 log("EA (RA) regval name", name
, regval
)
2474 inputs
[name
] = regval
2476 regval
= (yield from self
.get_input(name
, ew_src
, xlen
))
2477 log("regval name", name
, regval
)
2478 inputs
[name
] = regval
2480 # arrrrgh, awful hack, to get _RT into namespace
2481 if ins_name
in ['setvl', 'svstep']:
2483 RT
= yield self
.dec2
.dec
.RT
2484 self
.namespace
[regname
] = SelectableInt(RT
, 5)
2486 self
.namespace
["RT"] = SelectableInt(0, 5)
2487 regnum
, is_vec
= yield from get_idx_out(self
.dec2
, "RT")
2488 log('hack input reg %s %s' % (name
, str(regnum
)), is_vec
)
2490 # in SVP64 mode for LD/ST work out immediate
2491 # XXX TODO: replace_ds for DS-Form rather than D-Form.
2492 # use info.form to detect
2493 if self
.is_svp64_mode
and not ldst_imm_postinc
:
2494 yield from self
.check_replace_d(info
, remap_active
)
2496 # "special" registers
2497 for special
in info
.special_regs
:
2498 if special
in special_sprs
:
2499 inputs
[special
] = self
.spr
[special
]
2501 inputs
[special
] = self
.namespace
[special
]
2503 # clear trap (trap) NIA
2504 self
.trap_nia
= None
2506 # check if this was an sv.bc* and create an indicator that
2507 # this is the last check to be made as a loop. combined with
2508 # the ALL/ANY mode we can early-exit. note that BI (to test)
2509 # is an input so there is no termination if BI is scalar
2510 # (because early-termination is for *output* scalars)
2511 if self
.is_svp64_mode
and ins_name
.startswith("sv.bc"):
2512 end_loop
= srcstep
== vl
-1 or dststep
== vl
-1
2513 self
.namespace
['end_loop'] = SelectableInt(end_loop
, 1)
2515 inp_ca_ov
= (self
.spr
['XER'][XER_bits
['CA']].value
,
2516 self
.spr
['XER'][XER_bits
['OV']].value
)
2518 for k
, v
in inputs
.items():
2520 v
= SelectableInt(0, self
.XLEN
)
2521 # prevent pseudo-code from modifying input registers
2522 v
= copy_assign_rhs(v
)
2523 if isinstance(v
, SelectableInt
):
2527 # execute actual instruction here (finally)
2528 log("inputs", inputs
)
2529 inputs
= list(inputs
.values())
2530 results
= info
.func(self
, *inputs
)
2531 output_names
= create_args(info
.write_regs
)
2533 # record .ok before anything after the pseudo-code can modify it
2535 for out
, n
in zip(results
or [], output_names
):
2538 if isinstance(out
, SelectableInt
):
2540 log("results", outs
)
2541 log("results ok", outs_ok
)
2543 # "inject" decorator takes namespace from function locals: we need to
2544 # overwrite NIA being overwritten (sigh)
2545 if self
.trap_nia
is not None:
2546 self
.namespace
['NIA'] = self
.trap_nia
2548 log("after func", self
.namespace
['CIA'], self
.namespace
['NIA'])
2550 # check if op was a LD/ST so that debugging can check the
2552 if int_op
in [MicrOp
.OP_STORE
.value
,
2554 self
.last_st_addr
= self
.mem
.last_st_addr
2555 if int_op
in [MicrOp
.OP_LOAD
.value
,
2557 self
.last_ld_addr
= self
.mem
.last_ld_addr
2558 log("op", int_op
, MicrOp
.OP_STORE
.value
, MicrOp
.OP_LOAD
.value
,
2559 self
.last_st_addr
, self
.last_ld_addr
)
2561 # detect if CA/CA32 already in outputs (sra*, basically)
2563 ca32
= outs
.get("CA32")
2565 log("carry already done?", ca
, ca32
, output_names
)
2566 # soc test_pipe_caller tests don't have output_carry
2567 has_output_carry
= hasattr(self
.dec2
.e
.do
, "output_carry")
2568 carry_en
= has_output_carry
and (yield self
.dec2
.e
.do
.output_carry
)
2570 yield from self
.handle_carry_(
2571 inputs
, results
[0], ca
, ca32
, inp_ca_ov
=inp_ca_ov
)
2573 # get output named "overflow" and "CR0"
2574 overflow
= outs
.get('overflow')
2575 cr0
= outs
.get('CR0')
2576 cr1
= outs
.get('CR1')
2578 # soc test_pipe_caller tests don't have oe
2579 has_oe
= hasattr(self
.dec2
.e
.do
, "oe")
2580 # yeah just no. not in parallel processing
2581 if has_oe
and not self
.is_svp64_mode
:
2582 # detect if overflow was in return result
2583 ov_en
= yield self
.dec2
.e
.do
.oe
.oe
2584 ov_ok
= yield self
.dec2
.e
.do
.oe
.ok
2585 log("internal overflow", ins_name
, overflow
, "en?", ov_en
, ov_ok
)
2587 yield from self
.handle_overflow(
2588 inputs
, results
[0], overflow
, inp_ca_ov
=inp_ca_ov
)
2590 # only do SVP64 dest predicated Rc=1 if dest-pred is not enabled
2592 if not self
.is_svp64_mode
or not pred_dst_zero
:
2593 if hasattr(self
.dec2
.e
.do
, "rc"):
2594 rc_en
= yield self
.dec2
.e
.do
.rc
.rc
2595 # don't do Rc=1 for svstep it is handled explicitly.
2596 # XXX TODO: now that CR0 is supported, sort out svstep's pseudocode
2597 # to write directly to CR0 instead of in ISACaller. hooyahh.
2598 if rc_en
and ins_name
not in ['svstep']:
2599 if outs_ok
.get('FPSCR', False):
2600 FPSCR
= outs
['FPSCR']
2603 yield from self
.do_rc_ov(
2604 ins_name
, results
[0], overflow
, cr0
, cr1
, FPSCR
)
2607 ffirst_hit
= False, False
2608 if self
.is_svp64_mode
:
2609 sv_mode
= yield self
.dec2
.rm_dec
.sv_mode
2610 is_cr
= sv_mode
== SVMode
.CROP
.value
2611 chk
= rc_en
or is_cr
2612 if outs_ok
.get('CR', False):
2613 # early write so check_ffirst can see value
2614 self
.namespace
['CR'].eq(outs
['CR'])
2615 ffirst_hit
= (yield from self
.check_ffirst(info
, chk
, srcstep
))
2617 # any modified return results?
2618 yield from self
.do_outregs(
2619 info
, outs
, carry_en
, ffirst_hit
, ew_dst
, outs_ok
)
2621 # check if a FP Exception occurred. TODO for DD-FFirst, check VLi
2622 # and raise the exception *after* if VLi=1 but if VLi=0 then
2623 # truncate and make the exception "disappear".
2624 if self
.FPSCR
.FEX
and (self
.msr
[MSRb
.FE0
] or self
.msr
[MSRb
.FE1
]):
2625 self
.call_trap(0x700, PIb
.FP
)
2628 yield from self
.do_nia(asmop
, ins_name
, rc_en
, ffirst_hit
)
2631 def check_ffirst(self
, info
, rc_en
, srcstep
):
2632 """fail-first mode: checks a bit of Rc Vector, truncates VL
2634 rm_mode
= yield self
.dec2
.rm_dec
.mode
2635 ff_inv
= yield self
.dec2
.rm_dec
.inv
2636 cr_bit
= yield self
.dec2
.rm_dec
.cr_sel
2637 RC1
= yield self
.dec2
.rm_dec
.RC1
2638 vli_
= yield self
.dec2
.rm_dec
.vli
# VL inclusive if truncated
2639 log(" ff rm_mode", rc_en
, rm_mode
, SVP64RMMode
.FFIRST
.value
)
2643 log(" cr_bit", cr_bit
)
2644 log(" rc_en", rc_en
)
2645 ffirst
= yield from is_ffirst_mode(self
.dec2
)
2646 if not rc_en
or not ffirst
:
2648 # get the CR vevtor, do BO-test
2650 log("asmregs", info
.asmregs
[0], info
.write_regs
)
2651 if 'CR' in info
.write_regs
and 'BF' in info
.asmregs
[0]:
2653 regnum
, is_vec
= yield from get_cr_out(self
.dec2
, crf
)
2654 crtest
= self
.crl
[regnum
]
2655 ffirst_hit
= crtest
[cr_bit
] != ff_inv
2656 log("cr test", crf
, regnum
, int(crtest
), crtest
, cr_bit
, ff_inv
)
2657 log("cr test?", ffirst_hit
)
2660 # Fail-first activated, truncate VL
2661 vli
= SelectableInt(int(vli_
), 7)
2662 self
.svstate
.vl
= srcstep
+ vli
2663 yield self
.dec2
.state
.svstate
.eq(self
.svstate
.value
)
2664 yield Settle() # let decoder update
2667 def do_rc_ov(self
, ins_name
, result
, overflow
, cr0
, cr1
, FPSCR
):
2668 cr_out
= yield self
.dec2
.op
.cr_out
2669 if cr_out
== CROutSel
.CR1
.value
:
2673 regnum
, is_vec
= yield from get_cr_out(self
.dec2
, rc_reg
)
2674 # hang on... for `setvl` actually you want to test SVSTATE.VL
2675 is_setvl
= ins_name
in ('svstep', 'setvl')
2677 result
= SelectableInt(result
.vl
, 64)
2679 # overflow = None # do not override overflow except in setvl
2683 cr1
= int(FPSCR
.FX
) << 3
2684 cr1 |
= int(FPSCR
.FEX
) << 2
2685 cr1 |
= int(FPSCR
.VX
) << 1
2686 cr1 |
= int(FPSCR
.OX
)
2687 log("default fp cr1", cr1
)
2689 log("explicit cr1", cr1
)
2690 self
.crl
[regnum
].eq(cr1
)
2692 # if there was not an explicit CR0 in the pseudocode,
2694 c
= self
.handle_comparison(result
, regnum
, overflow
, no_so
=is_setvl
)
2695 log("implicit cr0 %d" % regnum
, c
)
2697 # otherwise we just blat CR0 into the required regnum
2698 log("explicit cr0 %d" % regnum
, cr0
)
2699 self
.crl
[regnum
].eq(cr0
)
2701 def do_outregs(self
, info
, outs
, ca_en
, ffirst_hit
, ew_dst
, outs_ok
):
2702 ffirst_hit
, vli
= ffirst_hit
2703 # write out any regs for this instruction, but only if fail-first is ok
2704 # XXX TODO: allow CR-vector to be written out even if ffirst fails
2705 if not ffirst_hit
or vli
:
2706 for name
, output
in outs
.items():
2707 if not outs_ok
[name
]:
2708 log("skipping writing output with .ok=False", name
, output
)
2710 yield from self
.check_write(info
, name
, output
, ca_en
, ew_dst
)
2711 # restore the CR value on non-VLI failfirst (from sv.cmp and others
2712 # which write directly to CR in the pseudocode (gah, what a mess)
2713 # if ffirst_hit and not vli:
2714 # self.cr.value = self.cr_backup
2716 def do_nia(self
, asmop
, ins_name
, rc_en
, ffirst_hit
):
2717 ffirst_hit
, vli
= ffirst_hit
2719 self
.svp64_reset_loop()
2722 # check advancement of src/dst/sub-steps and if PC needs updating
2723 nia_update
= (yield from self
.check_step_increment(
2724 rc_en
, asmop
, ins_name
))
2726 self
.update_pc_next()
2728 def check_replace_d(self
, info
, remap_active
):
2729 replace_d
= False # update / replace constant in pseudocode
2730 ldstmode
= yield self
.dec2
.rm_dec
.ldstmode
2731 vl
= self
.svstate
.vl
2732 subvl
= yield self
.dec2
.rm_dec
.rm_in
.subvl
2733 srcstep
, dststep
= self
.new_srcstep
, self
.new_dststep
2734 ssubstep
, dsubstep
= self
.new_ssubstep
, self
.new_dsubstep
2735 if info
.form
== 'DS':
2736 # DS-Form, multiply by 4 then knock 2 bits off after
2737 imm
= yield self
.dec2
.dec
.fields
.FormDS
.DS
[0:14] * 4
2739 imm
= yield self
.dec2
.dec
.fields
.FormD
.D
[0:16]
2740 imm
= exts(imm
, 16) # sign-extend to integer
2741 # get the right step. LD is from srcstep, ST is dststep
2742 op
= yield self
.dec2
.e
.do
.insn_type
2744 if op
== MicrOp
.OP_LOAD
.value
:
2746 offsmul
= yield self
.dec2
.in1_step
2747 log("D-field REMAP src", imm
, offsmul
, ldstmode
)
2749 offsmul
= (srcstep
* (subvl
+1)) + ssubstep
2750 log("D-field src", imm
, offsmul
, ldstmode
)
2751 elif op
== MicrOp
.OP_STORE
.value
:
2752 # XXX NOTE! no bit-reversed STORE! this should not ever be used
2753 offsmul
= (dststep
* (subvl
+1)) + dsubstep
2754 log("D-field dst", imm
, offsmul
, ldstmode
)
2755 # Unit-Strided LD/ST adds offset*width to immediate
2756 if ldstmode
== SVP64LDSTmode
.UNITSTRIDE
.value
:
2757 ldst_len
= yield self
.dec2
.e
.do
.data_len
2758 imm
= SelectableInt(imm
+ offsmul
* ldst_len
, 32)
2760 # Element-strided multiplies the immediate by element step
2761 elif ldstmode
== SVP64LDSTmode
.ELSTRIDE
.value
:
2762 imm
= SelectableInt(imm
* offsmul
, 32)
2765 ldst_ra_vec
= yield self
.dec2
.rm_dec
.ldst_ra_vec
2766 ldst_imz_in
= yield self
.dec2
.rm_dec
.ldst_imz_in
2767 log("LDSTmode", SVP64LDSTmode(ldstmode
),
2768 offsmul
, imm
, ldst_ra_vec
, ldst_imz_in
)
2769 # new replacement D... errr.. DS
2771 if info
.form
== 'DS':
2772 # TODO: assert 2 LSBs are zero?
2773 log("DS-Form, TODO, assert 2 LSBs zero?", bin(imm
.value
))
2774 imm
.value
= imm
.value
>> 2
2775 self
.namespace
['DS'] = imm
2777 self
.namespace
['D'] = imm
2779 def get_input(self
, name
, ew_src
, xlen
):
2780 # using PowerDecoder2, first, find the decoder index.
2781 # (mapping name RA RB RC RS to in1, in2, in3)
2782 regnum
, is_vec
= yield from get_idx_in(self
.dec2
, name
, True)
2784 # doing this is not part of svp64, it's because output
2785 # registers, to be modified, need to be in the namespace.
2786 regnum
, is_vec
= yield from get_idx_out(self
.dec2
, name
, True)
2788 regnum
, is_vec
= yield from get_idx_out2(self
.dec2
, name
, True)
2790 if isinstance(regnum
, tuple):
2791 (regnum
, base
, offs
) = regnum
2793 base
, offs
= regnum
, 0 # temporary HACK
2795 # in case getting the register number is needed, _RA, _RB
2796 # (HACK: only in straight non-svp64-mode for now, or elwidth == 64)
2797 regname
= "_" + name
2798 if not self
.is_svp64_mode
or ew_src
== 64:
2799 self
.namespace
[regname
] = regnum
2801 # FIXME: we're trying to access a sub-register, plain register
2802 # numbers don't work for that. for now, just pass something that
2803 # can be compared to 0 and probably will cause an error if misused.
2804 # see https://bugs.libre-soc.org/show_bug.cgi?id=1221
2805 self
.namespace
[regname
] = regnum
* 10000
2807 if not self
.is_svp64_mode
or not self
.pred_src_zero
:
2808 log('reading reg %s %s' % (name
, str(regnum
)), is_vec
)
2810 fval
= self
.fpr(base
, is_vec
, offs
, ew_src
)
2811 reg_val
= SelectableInt(fval
)
2812 assert ew_src
== self
.XLEN
, "TODO fix elwidth conversion"
2813 self
.trace("r:FPR:%d:%d:%d " % (base
, offs
, ew_src
))
2814 log("read fp reg %d/%d: 0x%x" % (base
, offs
, reg_val
.value
),
2815 kind
=LogType
.InstrInOuts
)
2816 elif name
is not None:
2817 gval
= self
.gpr(base
, is_vec
, offs
, ew_src
)
2818 reg_val
= SelectableInt(gval
.value
, bits
=xlen
)
2819 self
.trace("r:GPR:%d:%d:%d " % (base
, offs
, ew_src
))
2820 log("read int reg %d/%d: 0x%x" % (base
, offs
, reg_val
.value
),
2821 kind
=LogType
.InstrInOuts
)
2823 log('zero input reg %s %s' % (name
, str(regnum
)), is_vec
)
2824 reg_val
= SelectableInt(0, ew_src
)
2827 def remap_set_steps(self
, remaps
):
2828 """remap_set_steps sets up the in1/2/3 and out1/2 steps.
2829 they work in concert with PowerDecoder2 at the moment,
2830 there is no HDL implementation of REMAP. therefore this
2831 function, because ISACaller still uses PowerDecoder2,
2832 will *explicitly* write the dec2.XX_step values. this has
2835 # just some convenient debug info
2837 sname
= 'SVSHAPE%d' % i
2838 shape
= self
.spr
[sname
]
2839 log(sname
, bin(shape
.value
))
2840 log(" lims", shape
.lims
)
2841 log(" mode", shape
.mode
)
2842 log(" skip", shape
.skip
)
2844 # set up the list of steps to remap
2845 mi0
= self
.svstate
.mi0
2846 mi1
= self
.svstate
.mi1
2847 mi2
= self
.svstate
.mi2
2848 mo0
= self
.svstate
.mo0
2849 mo1
= self
.svstate
.mo1
2850 steps
= [[self
.dec2
.in1_step
, mi0
], # RA
2851 [self
.dec2
.in2_step
, mi1
], # RB
2852 [self
.dec2
.in3_step
, mi2
], # RC
2853 [self
.dec2
.o_step
, mo0
], # RT
2854 [self
.dec2
.o2_step
, mo1
], # EA
2857 rnames
= ['RA', 'RB', 'RC', 'RT', 'RS']
2858 for i
, reg
in enumerate(rnames
):
2859 idx
= yield from get_idx_map(self
.dec2
, reg
)
2861 idx
= yield from get_idx_map(self
.dec2
, "F"+reg
)
2863 steps
[i
][0] = self
.dec2
.in1_step
2865 steps
[i
][0] = self
.dec2
.in2_step
2867 steps
[i
][0] = self
.dec2
.in3_step
2868 log("remap step", i
, reg
, idx
, steps
[i
][1])
2869 remap_idxs
= self
.remap_idxs
2871 # now cross-index the required SHAPE for each of 3-in 2-out regs
2872 rnames
= ['RA', 'RB', 'RC', 'RT', 'EA']
2873 for i
, (dstep
, shape_idx
) in enumerate(steps
):
2874 (shape
, remap
) = remaps
[shape_idx
]
2875 remap_idx
= remap_idxs
[shape_idx
]
2876 # zero is "disabled"
2877 if shape
.value
== 0x0:
2879 # now set the actual requested step to the current index
2880 if dstep
is not None:
2881 yield dstep
.eq(remap_idx
)
2883 # debug printout info
2884 rremaps
.append((shape
.mode
, hex(shape
.value
), dstep
,
2885 i
, rnames
[i
], shape_idx
, remap_idx
))
2887 log("shape remap", x
)
2889 def check_write(self
, info
, name
, output
, carry_en
, ew_dst
):
2890 if name
== 'overflow': # ignore, done already (above)
2892 if name
== 'CR0': # ignore, done already (above)
2894 if isinstance(output
, int):
2895 output
= SelectableInt(output
, EFFECTIVELY_UNLIMITED
)
2897 if name
.startswith("RESERVE"):
2898 log("write %s 0x%x" % (name
, output
.value
))
2899 getattr(self
, name
).eq(output
)
2901 if name
in ['FPSCR', ]:
2902 log("write FPSCR 0x%x" % (output
.value
))
2903 self
.FPSCR
.eq(output
)
2906 if name
in ['CA', 'CA32']:
2908 log("writing %s to XER" % name
, output
)
2909 log("write XER %s 0x%x" % (name
, output
.value
))
2910 self
.spr
['XER'][XER_bits
[name
]] = output
.value
2912 log("NOT writing %s to XER" % name
, output
)
2914 # write special SPRs
2915 if name
in info
.special_regs
:
2916 log('writing special %s' % name
, output
, special_sprs
)
2917 log("write reg %s 0x%x" % (name
, output
.value
),
2918 kind
=LogType
.InstrInOuts
)
2919 if name
in special_sprs
:
2920 self
.spr
[name
] = output
2922 self
.namespace
[name
].eq(output
)
2924 log('msr written', hex(self
.msr
.value
))
2926 # find out1/out2 PR/FPR
2927 regnum
, is_vec
= yield from get_idx_out(self
.dec2
, name
, True)
2929 regnum
, is_vec
= yield from get_idx_out2(self
.dec2
, name
, True)
2931 # temporary hack for not having 2nd output
2932 regnum
= yield getattr(self
.decoder
, name
)
2934 # convenient debug prefix
2939 # check zeroing due to predicate bit being zero
2940 if self
.is_svp64_mode
and self
.pred_dst_zero
:
2941 log('zeroing reg %s %s' % (str(regnum
), str(output
)), is_vec
)
2942 output
= SelectableInt(0, EFFECTIVELY_UNLIMITED
)
2943 log("write reg %s%s 0x%x ew %d" % (reg_prefix
, str(regnum
),
2944 output
.value
, ew_dst
),
2945 kind
=LogType
.InstrInOuts
)
2946 # zero-extend tov64 bit begore storing (should use EXT oh well)
2947 if output
.bits
> 64:
2948 output
= SelectableInt(output
.value
, 64)
2949 rnum
, base
, offset
= regnum
2951 self
.fpr
.write(regnum
, output
, is_vec
, ew_dst
)
2952 self
.trace("w:FPR:%d:%d:%d " % (rnum
, offset
, ew_dst
))
2955 # LDST/Update does *not* allow elwidths on RA (Effective Address).
2956 # this has to be detected, and overridden. see get_input (related)
2957 sv_mode
= yield self
.dec2
.rm_dec
.sv_mode
2958 is_ldst
= (sv_mode
in [SVMode
.LDST_IDX
.value
, SVMode
.LDST_IMM
.value
] \
2959 and self
.is_svp64_mode
)
2960 if is_ldst
and name
in ['EA', 'RA']:
2961 op
= self
.dec2
.dec
.op
2962 if hasattr(op
, "upd"):
2963 # update mode LD/ST uses read-reg A also as an output
2965 log("write is_ldst is_update", sv_mode
, is_ldst
, upd
)
2966 if upd
== LDSTMode
.update
.value
:
2967 ew_dst
= 64 # override for RA (EA) to 64-bit
2969 self
.gpr
.write(regnum
, output
, is_vec
, ew_dst
)
2970 self
.trace("w:GPR:%d:%d:%d " % (rnum
, offset
, ew_dst
))
2972 def check_step_increment(self
, rc_en
, asmop
, ins_name
):
2973 # check if it is the SVSTATE.src/dest step that needs incrementing
2974 # this is our Sub-Program-Counter loop from 0 to VL-1
2975 if not self
.allow_next_step_inc
:
2976 if self
.is_svp64_mode
:
2977 return (yield from self
.svstate_post_inc(ins_name
))
2979 # XXX only in non-SVP64 mode!
2980 # record state of whether the current operation was an svshape,
2982 # to be able to know if it should apply in the next instruction.
2983 # also (if going to use this instruction) should disable ability
2984 # to interrupt in between. sigh.
2985 self
.last_op_svshape
= asmop
in ['svremap', 'svindex',
2992 log("SVSTATE_NEXT: inc requested, mode",
2993 self
.svstate_next_mode
, self
.allow_next_step_inc
)
2994 yield from self
.svstate_pre_inc()
2995 pre
= yield from self
.update_new_svstate_steps()
2997 # reset at end of loop including exit Vertical Mode
2998 log("SVSTATE_NEXT: end of loop, reset")
2999 self
.svp64_reset_loop()
3000 self
.svstate
.vfirst
= 0
3004 self
.handle_comparison(SelectableInt(0, 64)) # CR0
3006 if self
.allow_next_step_inc
== 2:
3007 log("SVSTATE_NEXT: read")
3008 nia_update
= (yield from self
.svstate_post_inc(ins_name
))
3010 log("SVSTATE_NEXT: post-inc")
3011 # use actual (cached) src/dst-step here to check end
3012 remaps
= self
.get_remap_indices()
3013 remap_idxs
= self
.remap_idxs
3014 vl
= self
.svstate
.vl
3015 subvl
= yield self
.dec2
.rm_dec
.rm_in
.subvl
3016 if self
.allow_next_step_inc
!= 2:
3017 yield from self
.advance_svstate_steps()
3018 #self.namespace['SVSTATE'] = self.svstate.spr
3019 # set CR0 (if Rc=1) based on end
3020 endtest
= 1 if self
.at_loopend() else 0
3022 #results = [SelectableInt(endtest, 64)]
3023 # self.handle_comparison(results) # CR0
3025 # see if svstep was requested, if so, which SVSTATE
3027 if self
.svstate_next_mode
> 0:
3028 shape_idx
= self
.svstate_next_mode
.value
-1
3029 endings
= self
.remap_loopends
[shape_idx
]
3030 cr_field
= SelectableInt((~endings
) << 1 | endtest
, 4)
3031 log("svstep Rc=1, CR0", cr_field
, endtest
)
3032 self
.crl
[0].eq(cr_field
) # CR0
3034 # reset at end of loop including exit Vertical Mode
3035 log("SVSTATE_NEXT: after increments, reset")
3036 self
.svp64_reset_loop()
3037 self
.svstate
.vfirst
= 0
3040 def SVSTATE_NEXT(self
, mode
, submode
, RA
=None):
3041 """explicitly moves srcstep/dststep on to next element, for
3042 "Vertical-First" mode. this function is called from
3043 setvl pseudo-code, as a pseudo-op "svstep"
3045 WARNING: this function uses information that was created EARLIER
3046 due to it being in the middle of a yield, but this function is
3047 *NOT* called from yield (it's called from compiled pseudocode).
3049 self
.allow_next_step_inc
= submode
.value
+ 1
3050 log("SVSTATE_NEXT mode", mode
, submode
, self
.allow_next_step_inc
)
3051 self
.svstate_next_mode
= mode
3052 if self
.svstate_next_mode
> 0 and self
.svstate_next_mode
< 5:
3053 shape_idx
= self
.svstate_next_mode
.value
-1
3054 return SelectableInt(self
.remap_idxs
[shape_idx
], 7)
3055 if self
.svstate_next_mode
== 5:
3056 self
.svstate_next_mode
= 0
3057 return SelectableInt(self
.svstate
.srcstep
, 7)
3058 if self
.svstate_next_mode
== 6:
3059 self
.svstate_next_mode
= 0
3060 return SelectableInt(self
.svstate
.dststep
, 7)
3061 if self
.svstate_next_mode
== 7:
3062 self
.svstate_next_mode
= 0
3063 return SelectableInt(self
.svstate
.ssubstep
, 7)
3064 if self
.svstate_next_mode
== 8:
3065 self
.svstate_next_mode
= 0
3066 return SelectableInt(self
.svstate
.dsubstep
, 7)
3067 return SelectableInt(0, 7)
3069 def get_src_dststeps(self
):
3070 """gets srcstep, dststep, and ssubstep, dsubstep
3072 return (self
.new_srcstep
, self
.new_dststep
,
3073 self
.new_ssubstep
, self
.new_dsubstep
)
3075 def update_svstate_namespace(self
, overwrite_svstate
=True):
3076 if overwrite_svstate
:
3077 # note, do not get the bit-reversed srcstep here!
3078 srcstep
, dststep
= self
.new_srcstep
, self
.new_dststep
3079 ssubstep
, dsubstep
= self
.new_ssubstep
, self
.new_dsubstep
3081 # update SVSTATE with new srcstep
3082 self
.svstate
.srcstep
= srcstep
3083 self
.svstate
.dststep
= dststep
3084 self
.svstate
.ssubstep
= ssubstep
3085 self
.svstate
.dsubstep
= dsubstep
3086 self
.namespace
['SVSTATE'] = self
.svstate
3087 yield self
.dec2
.state
.svstate
.eq(self
.svstate
.value
)
3088 yield Settle() # let decoder update
3090 def update_new_svstate_steps(self
, overwrite_svstate
=True):
3091 yield from self
.update_svstate_namespace(overwrite_svstate
)
3092 srcstep
= self
.svstate
.srcstep
3093 dststep
= self
.svstate
.dststep
3094 ssubstep
= self
.svstate
.ssubstep
3095 dsubstep
= self
.svstate
.dsubstep
3096 pack
= self
.svstate
.pack
3097 unpack
= self
.svstate
.unpack
3098 vl
= self
.svstate
.vl
3099 sv_mode
= yield self
.dec2
.rm_dec
.sv_mode
3100 subvl
= yield self
.dec2
.rm_dec
.rm_in
.subvl
3101 rm_mode
= yield self
.dec2
.rm_dec
.mode
3102 ff_inv
= yield self
.dec2
.rm_dec
.inv
3103 cr_bit
= yield self
.dec2
.rm_dec
.cr_sel
3104 log(" srcstep", srcstep
)
3105 log(" dststep", dststep
)
3107 log(" unpack", unpack
)
3108 log(" ssubstep", ssubstep
)
3109 log(" dsubstep", dsubstep
)
3111 log(" subvl", subvl
)
3112 log(" rm_mode", rm_mode
)
3113 log(" sv_mode", sv_mode
)
3115 log(" cr_bit", cr_bit
)
3117 # check if end reached (we let srcstep overrun, above)
3118 # nothing needs doing (TODO zeroing): just do next instruction
3121 return ((ssubstep
== subvl
and srcstep
== vl
) or
3122 (dsubstep
== subvl
and dststep
== vl
))
3124 def svstate_post_inc(self
, insn_name
, vf
=0):
3125 # check if SV "Vertical First" mode is enabled
3126 vfirst
= self
.svstate
.vfirst
3127 log(" SV Vertical First", vf
, vfirst
)
3128 if not vf
and vfirst
== 1:
3129 # SV Branch-Conditional required to be as-if-vector
3130 # because there *is* no destination register
3131 # (SV normally only terminates on 1st scalar reg written
3132 # except in [slightly-misnamed] mapreduce mode)
3133 ffirst
= yield from is_ffirst_mode(self
.dec2
)
3134 if insn_name
.startswith("sv.bc") or ffirst
:
3135 self
.update_pc_next()
3140 # check if it is the SVSTATE.src/dest step that needs incrementing
3141 # this is our Sub-Program-Counter loop from 0 to VL-1
3142 # XXX twin predication TODO
3143 vl
= self
.svstate
.vl
3144 subvl
= yield self
.dec2
.rm_dec
.rm_in
.subvl
3145 mvl
= self
.svstate
.maxvl
3146 srcstep
= self
.svstate
.srcstep
3147 dststep
= self
.svstate
.dststep
3148 ssubstep
= self
.svstate
.ssubstep
3149 dsubstep
= self
.svstate
.dsubstep
3150 pack
= self
.svstate
.pack
3151 unpack
= self
.svstate
.unpack
3152 rm_mode
= yield self
.dec2
.rm_dec
.mode
3153 reverse_gear
= yield self
.dec2
.rm_dec
.reverse_gear
3154 sv_ptype
= yield self
.dec2
.dec
.op
.SV_Ptype
3155 out_vec
= not (yield self
.dec2
.no_out_vec
)
3156 in_vec
= not (yield self
.dec2
.no_in_vec
)
3157 rm_mode
= yield self
.dec2
.rm_dec
.mode
3158 log(" svstate.vl", vl
)
3159 log(" svstate.mvl", mvl
)
3160 log(" rm.subvl", subvl
)
3161 log(" svstate.srcstep", srcstep
)
3162 log(" svstate.dststep", dststep
)
3163 log(" svstate.ssubstep", ssubstep
)
3164 log(" svstate.dsubstep", dsubstep
)
3165 log(" svstate.pack", pack
)
3166 log(" svstate.unpack", unpack
)
3167 log(" mode", rm_mode
)
3168 log(" reverse", reverse_gear
)
3169 log(" out_vec", out_vec
)
3170 log(" in_vec", in_vec
)
3171 log(" sv_ptype", sv_ptype
, sv_ptype
== SVPType
.P2
.value
)
3172 log(" rm_mode", rm_mode
)
3173 # check if this was an sv.bc* and if so did it succeed
3174 if self
.is_svp64_mode
and insn_name
.startswith("sv.bc"):
3175 end_loop
= self
.namespace
['end_loop']
3176 log("branch %s end_loop" % insn_name
, end_loop
)
3178 self
.svp64_reset_loop()
3179 self
.update_pc_next()
3181 # check if srcstep needs incrementing by one, stop PC advancing
3182 # but for 2-pred both src/dest have to be checked.
3183 # XXX this might not be true! it may just be LD/ST
3184 if sv_ptype
== SVPType
.P2
.value
:
3185 svp64_is_vector
= (out_vec
or in_vec
)
3187 svp64_is_vector
= out_vec
3188 # also if data-dependent fail-first is used, only in_vec is tested,
3189 # allowing *scalar destinations* to be used as an accumulator.
3190 # effectively this implies /mr (mapreduce mode) is 100% on with ddffirst
3191 # see https://bugs.libre-soc.org/show_bug.cgi?id=1183#c16
3192 ffirst
= yield from is_ffirst_mode(self
.dec2
)
3194 svp64_is_vector
= in_vec
3196 # loops end at the first "hit" (source or dest)
3197 yield from self
.advance_svstate_steps()
3198 loopend
= self
.loopend
3199 log("loopend", svp64_is_vector
, loopend
)
3200 if not svp64_is_vector
or loopend
:
3201 # reset loop to zero and update NIA
3202 self
.svp64_reset_loop()
3207 # still looping, advance and update NIA
3208 self
.namespace
['SVSTATE'] = self
.svstate
3210 # not an SVP64 branch, so fix PC (NIA==CIA) for next loop
3211 # (by default, NIA is CIA+4 if v3.0B or CIA+8 if SVP64)
3212 # this way we keep repeating the same instruction (with new steps)
3213 self
.pc
.NIA
.eq(self
.pc
.CIA
)
3214 self
.namespace
['NIA'] = self
.pc
.NIA
3215 log("end of sub-pc call", self
.namespace
['CIA'], self
.namespace
['NIA'])
3216 return False # DO NOT allow PC update whilst Sub-PC loop running
3218 def update_pc_next(self
):
3219 # UPDATE program counter
3220 self
.pc
.update(self
.namespace
, self
.is_svp64_mode
)
3221 #self.svstate.spr = self.namespace['SVSTATE']
3222 log("end of call", self
.namespace
['CIA'],
3223 self
.namespace
['NIA'],
3224 self
.namespace
['SVSTATE'])
3226 def svp64_reset_loop(self
):
3227 self
.svstate
.srcstep
= 0
3228 self
.svstate
.dststep
= 0
3229 self
.svstate
.ssubstep
= 0
3230 self
.svstate
.dsubstep
= 0
3231 self
.loopend
= False
3232 log(" svstate.srcstep loop end (PC to update)")
3233 self
.namespace
['SVSTATE'] = self
.svstate
3235 def update_nia(self
):
3236 self
.pc
.update_nia(self
.is_svp64_mode
)
3237 self
.namespace
['NIA'] = self
.pc
.NIA
3241 """Decorator factory.
3243 this decorator will "inject" variables into the function's namespace,
3244 from the *dictionary* in self.namespace. it therefore becomes possible
3245 to make it look like a whole stack of variables which would otherwise
3246 need "self." inserted in front of them (*and* for those variables to be
3247 added to the instance) "appear" in the function.
3249 "self.namespace['SI']" for example becomes accessible as just "SI" but
3250 *only* inside the function, when decorated.
3252 def variable_injector(func
):
3254 def decorator(*args
, **kwargs
):
3256 func_globals
= func
.__globals
__ # Python 2.6+
3257 except AttributeError:
3258 func_globals
= func
.func_globals
# Earlier versions.
3260 context
= args
[0].namespace
# variables to be injected
3261 saved_values
= func_globals
.copy() # Shallow copy of dict.
3262 log("globals before", context
.keys())
3263 func_globals
.update(context
)
3264 result
= func(*args
, **kwargs
)
3265 log("globals after", func_globals
['CIA'], func_globals
['NIA'])
3266 log("args[0]", args
[0].namespace
['CIA'],
3267 args
[0].namespace
['NIA'],
3268 args
[0].namespace
['SVSTATE'])
3269 if 'end_loop' in func_globals
:
3270 log("args[0] end_loop", func_globals
['end_loop'])
3271 args
[0].namespace
= func_globals
3272 #exec (func.__code__, func_globals)
3275 # func_globals = saved_values # Undo changes.
3281 return variable_injector