1 # SPDX-License-Identifier: LGPLv3+
2 # Copyright (C) 2021 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
3 # Funded by NLnet http://nlnet.nl
5 """SVP64 OpenPOWER v3.0B assembly translator
7 This class takes raw svp64 assembly mnemonics (aliases excluded) and
8 creates an EXT001-encoded "svp64 prefix" followed by a v3.0B opcode.
10 It is very simple and straightforward, the only weirdness being the
11 extraction of the register information and conversion to v3.0B numbering.
13 Encoding format of svp64: https://libre-soc.org/openpower/sv/svp64/
14 Bugtracker: https://bugs.libre-soc.org/show_bug.cgi?id=578
18 from collections
import OrderedDict
20 from soc
.decoder
.isa
.caller
import (SVP64PrefixFields
, SV64P_MAJOR_SIZE
,
21 SV64P_PID_SIZE
, SVP64RMFields
,
22 SVP64RM_EXTRA2_SPEC_SIZE
,
23 SVP64RM_EXTRA3_SPEC_SIZE
,
24 SVP64RM_MODE_SIZE
, SVP64RM_SMASK_SIZE
,
25 SVP64RM_MMODE_SIZE
, SVP64RM_MASK_SIZE
,
26 SVP64RM_SUBVL_SIZE
, SVP64RM_EWSRC_SIZE
,
28 from soc
.decoder
.pseudo
.pagereader
import ISA
29 from soc
.decoder
.power_svp64
import SVP64RM
, get_regtype
, decode_extra
30 from soc
.decoder
.selectable_int
import SelectableInt
33 # decode GPR into sv extra
34 def get_extra_gpr(etype
, regmode
, field
):
35 if regmode
== 'scalar':
36 # cut into 2-bits 5-bits SS FFFFF
38 field
= field
& 0b11111
40 # cut into 5-bits 2-bits FFFFF SS
41 sv_extra
= field
& 0b11
43 return sv_extra
, field
46 # decode 3-bit CR into sv extra
47 def get_extra_cr_3bit(etype
, regmode
, field
):
48 if regmode
== 'scalar':
49 # cut into 2-bits 3-bits SS FFF
53 # cut into 3-bits 4-bits FFF SSSS but will cut 2 zeros off later
54 sv_extra
= field
& 0b1111
56 return sv_extra
, field
60 def decode_subvl(encoding
):
61 pmap
= {'2': 0b01, '3': 0b10, '4': 0b11}
62 assert encoding
in pmap
, \
63 "encoding %s for SUBVL not recognised" % encoding
68 def decode_elwidth(encoding
):
69 pmap
= {'8': 0b11, '16': 0b10, '32': 0b01}
70 assert encoding
in pmap
, \
71 "encoding %s for elwidth not recognised" % encoding
75 # decodes predicate register encoding
76 def decode_predicate(encoding
):
87 'nl' : (1, 0b001), 'ge' : (1, 0b001), # same value
89 'ng' : (1, 0b011), 'le' : (1, 0b011), # same value
92 'so' : (1, 0b110), 'un' : (1, 0b110), # same value
93 'ns' : (1, 0b111), 'nu' : (1, 0b111), # same value
95 assert encoding
in pmap
, \
96 "encoding %s for predicate not recognised" % encoding
100 # decodes "Mode" in similar way to BO field (supposed to, anyway)
101 def decode_bo(encoding
):
102 pmap
= { # TODO: double-check that these are the same as Branch BO
104 'nl' : 0b001, 'ge' : 0b001, # same value
106 'ng' : 0b011, 'le' : 0b011, # same value
109 'so' : 0b110, 'un' : 0b110, # same value
110 'ns' : 0b111, 'nu' : 0b111, # same value
112 assert encoding
in pmap
, \
113 "encoding %s for BO Mode not recognised" % encoding
114 return pmap
[encoding
]
116 # partial-decode fail-first mode
117 def decode_ffirst(encoding
):
118 if encoding
in ['RC1', '~RC1']:
120 return decode_bo(encoding
)
123 def decode_reg(field
):
124 # decode the field number. "5.v" or "3.s" or "9"
125 field
= field
.split(".")
126 regmode
= 'scalar' # default
130 elif field
[1] == 'v':
132 field
= int(field
[0]) # actual register number
133 return field
, regmode
136 def decode_imm(field
):
137 ldst_imm
= "(" in field
and field
[-1] == ')'
139 return field
[:-1].split("(")
143 # decodes svp64 assembly listings and creates EXT001 svp64 prefixes
145 def __init__(self
, lst
):
147 self
.trans
= self
.translate(lst
)
150 yield from self
.trans
152 def translate(self
, lst
):
153 isa
= ISA() # reads the v3.0B pseudo-code markdown files
154 svp64
= SVP64RM() # reads the svp64 Remap entries for registers
156 # find first space, to get opcode
159 # now find opcode fields
160 fields
= ''.join(ls
[1:]).split(',')
161 fields
= list(map(str.strip
, fields
))
162 print ("opcode, fields", ls
, opcode
, fields
)
164 # identify if is a svp64 mnemonic
165 if not opcode
.startswith('sv.'):
166 yield insn
# unaltered
168 opcode
= opcode
[3:] # strip leading "sv."
170 # start working on decoding the svp64 op: sv.basev30Bop/vec2/mode
171 opmodes
= opcode
.split("/") # split at "/"
172 v30b_op
= opmodes
.pop(0) # first is the v3.0B
173 # check instruction ends with dot
174 rc_mode
= v30b_op
.endswith('.')
176 v30b_op
= v30b_op
[:-1]
178 if v30b_op
not in isa
.instr
:
179 raise Exception("opcode %s of '%s' not supported" % \
181 if v30b_op
not in svp64
.instrs
:
182 raise Exception("opcode %s of '%s' not an svp64 instruction" % \
184 v30b_regs
= isa
.instr
[v30b_op
].regs
[0] # get regs info "RT, RA, RB"
185 rm
= svp64
.instrs
[v30b_op
] # one row of the svp64 RM CSV
186 print ("v3.0B op", v30b_op
, "Rc=1" if rc_mode
else '')
187 print ("v3.0B regs", opcode
, v30b_regs
)
190 # right. the first thing to do is identify the ordering of
191 # the registers, by name. the EXTRA2/3 ordering is in
192 # rm['0']..rm['3'] but those fields contain the names RA, BB
193 # etc. we have to read the pseudocode to understand which
194 # reg is which in our instruction. sigh.
196 # first turn the svp64 rm into a "by name" dict, recording
197 # which position in the RM EXTRA it goes into
198 # also: record if the src or dest was a CR, for sanity-checking
199 # (elwidth overrides on CRs are banned)
200 decode
= decode_extra(rm
)
201 dest_reg_cr
, src_reg_cr
, svp64_src
, svp64_dest
= decode
202 svp64_reg_byname
= {}
203 svp64_reg_byname
.update(svp64_src
)
204 svp64_reg_byname
.update(svp64_dest
)
206 print ("EXTRA field index, by regname", svp64_reg_byname
)
208 # okaaay now we identify the field value (opcode N,N,N) with
209 # the pseudo-code info (opcode RT, RA, RB)
210 assert len(fields
) == len(v30b_regs
), \
211 "length of fields %s must match insn `%s`" % \
212 (str(v30b_regs
), insn
)
213 opregfields
= zip(fields
, v30b_regs
) # err that was easy
215 # now for each of those find its place in the EXTRA encoding
216 extras
= OrderedDict()
217 for idx
, (field
, regname
) in enumerate(opregfields
):
218 extra
= svp64_reg_byname
.get(regname
, None)
219 imm
, regname
= decode_imm(regname
)
220 rtype
= get_regtype(regname
)
221 extras
[extra
] = (idx
, field
, regname
, rtype
, imm
)
222 print (" ", extra
, extras
[extra
])
224 # great! got the extra fields in their associated positions:
225 # also we know the register type. now to create the EXTRA encodings
226 etype
= rm
['Etype'] # Extra type: EXTRA3/EXTRA2
227 ptype
= rm
['Ptype'] # Predication type: Twin / Single
230 for extra_idx
, (idx
, field
, rname
, rtype
, iname
) in extras
.items():
231 # is it a field we don't alter/examine? if so just put it
234 v30b_newfields
.append(field
)
236 # identify if this is a ld/st immediate(reg) thing
237 ldst_imm
= "(" in field
and field
[-1] == ')'
239 immed
, field
= field
[:-1].split("(")
241 field
, regmode
= decode_reg(field
)
242 print (" ", rtype
, regmode
, iname
, field
, end
=" ")
244 # see Mode field https://libre-soc.org/openpower/sv/svp64/
245 # XXX TODO: the following is a bit of a laborious repeated
246 # mess, which could (and should) easily be parameterised.
247 # XXX also TODO: the LD/ST modes which are different
248 # https://libre-soc.org/openpower/sv/ldst/
250 # encode SV-GPR field into extra, v3.0field
252 sv_extra
, field
= get_extra_gpr(etype
, regmode
, field
)
253 # now sanity-check. EXTRA3 is ok, EXTRA2 has limits
254 # (and shrink to a single bit if ok)
255 if etype
== 'EXTRA2':
256 if regmode
== 'scalar':
257 # range is r0-r63 in increments of 1
258 assert (sv_extra
>> 1) == 0, \
259 "scalar GPR %s cannot fit into EXTRA2 %s" % \
260 (rname
, str(extras
[extra_idx
]))
261 # all good: encode as scalar
262 sv_extra
= sv_extra
& 0b01
264 # range is r0-r127 in increments of 4
265 assert sv_extra
& 0b01 == 0, \
266 "vector field %s cannot fit into EXTRA2 %s" % \
267 (rname
, str(extras
[extra_idx
]))
268 # all good: encode as vector (bit 2 set)
269 sv_extra
= 0b10 |
(sv_extra
>> 1)
270 elif regmode
== 'vector':
271 # EXTRA3 vector bit needs marking
274 # encode SV-CR 3-bit field into extra, v3.0field
275 elif rtype
== 'CR_3bit':
276 sv_extra
, field
= get_extra_cr_3bit(etype
, regmode
, field
)
277 # now sanity-check (and shrink afterwards)
278 if etype
== 'EXTRA2':
279 if regmode
== 'scalar':
280 # range is CR0-CR15 in increments of 1
281 assert (sv_extra
>> 1) == 0, \
282 "scalar CR %s cannot fit into EXTRA2 %s" % \
283 (rname
, str(extras
[extra_idx
]))
284 # all good: encode as scalar
285 sv_extra
= sv_extra
& 0b01
287 # range is CR0-CR127 in increments of 16
288 assert sv_extra
& 0b111 == 0, \
289 "vector CR %s cannot fit into EXTRA2 %s" % \
290 (rname
, str(extras
[extra_idx
]))
291 # all good: encode as vector (bit 2 set)
292 sv_extra
= 0b10 |
(sv_extra
>> 3)
294 if regmode
== 'scalar':
295 # range is CR0-CR31 in increments of 1
296 assert (sv_extra
>> 2) == 0, \
297 "scalar CR %s cannot fit into EXTRA2 %s" % \
298 (rname
, str(extras
[extra_idx
]))
299 # all good: encode as scalar
300 sv_extra
= sv_extra
& 0b11
302 # range is CR0-CR127 in increments of 8
303 assert sv_extra
& 0b11 == 0, \
304 "vector CR %s cannot fit into EXTRA2 %s" % \
305 (rname
, str(extras
[extra_idx
]))
306 # all good: encode as vector (bit 3 set)
307 sv_extra
= 0b100 |
(sv_extra
>> 2)
309 # encode SV-CR 5-bit field into extra, v3.0field
310 # *sigh* this is the same as 3-bit except the 2 LSBs are
312 elif rtype
== 'CR_5bit':
313 cr_subfield
= field
& 0b11
314 field
= field
>> 2 # strip bottom 2 bits
315 sv_extra
, field
= get_extra_cr_3bit(etype
, regmode
, field
)
316 # now sanity-check (and shrink afterwards)
317 if etype
== 'EXTRA2':
318 if regmode
== 'scalar':
319 # range is CR0-CR15 in increments of 1
320 assert (sv_extra
>> 1) == 0, \
321 "scalar CR %s cannot fit into EXTRA2 %s" % \
322 (rname
, str(extras
[extra_idx
]))
323 # all good: encode as scalar
324 sv_extra
= sv_extra
& 0b01
326 # range is CR0-CR127 in increments of 16
327 assert sv_extra
& 0b111 == 0, \
328 "vector CR %s cannot fit into EXTRA2 %s" % \
329 (rname
, str(extras
[extra_idx
]))
330 # all good: encode as vector (bit 2 set)
331 sv_extra
= 0b10 |
(sv_extra
>> 3)
333 if regmode
== 'scalar':
334 # range is CR0-CR31 in increments of 1
335 assert (sv_extra
>> 2) == 0, \
336 "scalar CR %s cannot fit into EXTRA2 %s" % \
337 (rname
, str(extras
[extra_idx
]))
338 # all good: encode as scalar
339 sv_extra
= sv_extra
& 0b11
341 # range is CR0-CR127 in increments of 8
342 assert sv_extra
& 0b11 == 0, \
343 "vector CR %s cannot fit into EXTRA2 %s" % \
344 (rname
, str(extras
[extra_idx
]))
345 # all good: encode as vector (bit 3 set)
346 sv_extra
= 0b100 |
(sv_extra
>> 2)
348 # reconstruct the actual 5-bit CR field
349 field
= (field
<< 2) | cr_subfield
351 # capture the extra field info
352 print ("=>", "%5s" % bin(sv_extra
), field
)
353 extras
[extra_idx
] = sv_extra
355 # append altered field value to v3.0b, differs for LDST
357 v30b_newfields
.append(("%s(%s)" % (immed
, str(field
))))
359 v30b_newfields
.append(str(field
))
361 print ("new v3.0B fields", v30b_op
, v30b_newfields
)
362 print ("extras", extras
)
364 # rright. now we have all the info. start creating SVP64 RM
365 svp64_rm
= SVP64RMFields()
367 # begin with EXTRA fields
368 for idx
, sv_extra
in extras
.items():
369 if idx
is None: continue
370 if etype
== 'EXTRA2':
371 svp64_rm
.extra2
[idx
].eq(
372 SelectableInt(sv_extra
, SVP64RM_EXTRA2_SPEC_SIZE
))
374 svp64_rm
.extra3
[idx
].eq(
375 SelectableInt(sv_extra
, SVP64RM_EXTRA3_SPEC_SIZE
))
380 destwid
= 0 # bits 4-5
381 srcwid
= 0 # bits 6-7
383 smask
= 0 # bits 16-18 but only for twin-predication
384 mode
= 0 # bits 19-23
395 mapreduce_crm
= False
396 mapreduce_svm
= False
401 # ok let's start identifying opcode augmentation fields
402 for encmode
in opmodes
:
403 # predicate mask (dest)
404 if encmode
.startswith("m="):
406 pmmode
, pmask
= decode_predicate(encmode
[2:])
409 # predicate mask (src, twin-pred)
410 elif encmode
.startswith("sm="):
412 smmode
, smask
= decode_predicate(encmode
[3:])
416 elif encmode
.startswith("vec"):
417 subvl
= decode_subvl(encmode
[3:])
419 elif encmode
.startswith("ew="):
420 destwid
= decode_elwidth(encmode
[3:])
421 elif encmode
.startswith("sw="):
422 srcwid
= decode_elwidth(encmode
[3:])
424 elif encmode
== 'sats':
425 assert sv_mode
is None
428 elif encmode
== 'satu':
429 assert sv_mode
is None
433 elif encmode
== 'sz':
435 elif encmode
== 'dz':
438 elif encmode
.startswith("ff="):
439 assert sv_mode
is None
441 failfirst
= decode_ffirst(encmode
[3:])
442 # predicate-result, interestingly same as fail-first
443 elif encmode
.startswith("pr="):
444 assert sv_mode
is None
446 predresult
= decode_ffirst(encmode
[3:])
448 elif encmode
== 'mr':
449 assert sv_mode
is None
452 elif encmode
== 'crm': # CR on map-reduce
453 assert sv_mode
is None
456 elif encmode
== 'svm': # sub-vector mode
459 # sanity-check that 2Pred mask is same mode
460 if has_pmask
and has_smask
:
461 assert smmode
== pmmode
, \
462 "predicate masks %s and %s must be same reg type" % \
465 # sanity-check that twin-predication mask only specified in 2P mode
467 assert has_smask
== False, \
468 "source-mask can only be specified on Twin-predicate ops"
470 # construct the mode field, doing sanity-checking along the way
473 assert sv_mode
== 0b00, "sub-vector mode in mapreduce only"
474 assert subvl
!= 0, "sub-vector mode not possible on SUBVL=1"
477 assert has_smask
, "src zeroing requires a source predicate"
479 assert has_pmask
, "dest zeroing requires a dest predicate"
483 mode |
= (src_zero
<< 3) |
(dst_zero
<< 4) # predicate zeroing
487 elif sv_mode
== 0b00:
488 mode |
= (0b1<<2) # sets mapreduce
489 assert dst_zero
== 0, "dest-zero not allowed in mapreduce mode"
491 mode |
= (0b1<<4) # sets CRM mode
492 assert rc_mode
, "CRM only allowed when Rc=1"
493 # bit of weird encoding to jam zero-pred or SVM mode in.
494 # SVM mode can be enabled only when SUBVL=2/3/4 (vec2/3/4)
496 mode |
= (src_zero
<< 3) # predicate src-zeroing
498 mode |
= (1 << 3) # SVM mode
501 elif sv_mode
== 0b01:
502 assert dst_zero
== 0, "dest-zero not allowed in failfirst mode"
503 if failfirst
== 'RC1':
504 mode |
= (0b1<<4) # sets RC1 mode
505 mode |
= (src_zero
<< 3) # predicate src-zeroing
506 assert rc_mode
==False, "ffirst RC1 only possible when Rc=0"
507 elif failfirst
== '~RC1':
508 mode |
= (0b1<<4) # sets RC1 mode...
509 mode |
= (src_zero
<< 3) # predicate src-zeroing
510 mode |
= (0b1<<2) # ... with inversion
511 assert rc_mode
==False, "ffirst RC1 only possible when Rc=0"
513 assert src_zero
== 0, "src-zero not allowed in ffirst BO"
514 assert rc_mode
, "ffirst BO only possible when Rc=1"
515 mode |
= (failfirst
<< 2) # set BO
518 elif sv_mode
== 0b10:
519 mode |
= (src_zero
<< 3) |
(dst_zero
<< 4) # predicate zeroing
520 mode |
= (saturation
<<2) # sets signed/unsigned saturation
522 # "predicate-result" modes. err... code-duplication from ffirst
523 elif sv_mode
== 0b11:
524 assert dst_zero
== 0, "dest-zero not allowed in predresult mode"
525 if predresult
== 'RC1':
526 mode |
= (0b1<<4) # sets RC1 mode
527 mode |
= (src_zero
<< 3) # predicate src-zeroing
528 assert rc_mode
==False, "pr-mode RC1 only possible when Rc=0"
529 elif predresult
== '~RC1':
530 mode |
= (0b1<<4) # sets RC1 mode...
531 mode |
= (src_zero
<< 3) # predicate src-zeroing
532 mode |
= (0b1<<2) # ... with inversion
533 assert rc_mode
==False, "pr-mode RC1 only possible when Rc=0"
535 assert src_zero
== 0, "src-zero not allowed in pr-mode BO"
536 assert rc_mode
, "pr-mode BO only possible when Rc=1"
537 mode |
= (predresult
<< 2) # set BO
539 # whewww.... modes all done :)
540 # now put into svp64_rm
543 svp64_rm
.mode
.eq(SelectableInt(mode
, SVP64RM_MODE_SIZE
))
545 # put in predicate masks into svp64_rm
547 # source pred: bits 16-18
548 svp64_rm
.smask
.eq(SelectableInt(smask
, SVP64RM_SMASK_SIZE
))
550 svp64_rm
.mmode
.eq(SelectableInt(mmode
, SVP64RM_MMODE_SIZE
))
552 svp64_rm
.mask
.eq(SelectableInt(pmask
, SVP64RM_MASK_SIZE
))
554 # and subvl: bits 8-9
555 svp64_rm
.subvl
.eq(SelectableInt(subvl
, SVP64RM_SUBVL_SIZE
))
559 svp64_rm
.ewsrc
.eq(SelectableInt(srcwid
, SVP64RM_EWSRC_SIZE
))
561 svp64_rm
.elwidth
.eq(SelectableInt(destwid
, SVP64RM_ELWIDTH_SIZE
))
563 # nice debug printout. (and now for something completely different)
564 # https://youtu.be/u0WOIwlXE9g?t=146
565 svp64_rm_value
= svp64_rm
.spr
.value
566 print ("svp64_rm", hex(svp64_rm_value
), bin(svp64_rm_value
))
567 print (" mmode 0 :", bin(mmode
))
568 print (" pmask 1-3 :", bin(pmask
))
569 print (" dstwid 4-5 :", bin(destwid
))
570 print (" srcwid 6-7 :", bin(srcwid
))
571 print (" subvl 8-9 :", bin(subvl
))
572 print (" mode 19-23:", bin(mode
))
573 offs
= 2 if etype
== 'EXTRA2' else 3 # 2 or 3 bits
574 for idx
, sv_extra
in extras
.items():
575 if idx
is None: continue
576 start
= (10+idx
*offs
)
578 print (" extra%d %2d-%2d:" % (idx
, start
, end
),
581 print (" smask 16-17:", bin(smask
))
584 # first, construct the prefix from its subfields
585 svp64_prefix
= SVP64PrefixFields()
586 svp64_prefix
.major
.eq(SelectableInt(0x1, SV64P_MAJOR_SIZE
))
587 svp64_prefix
.pid
.eq(SelectableInt(0b11, SV64P_PID_SIZE
))
588 svp64_prefix
.rm
.eq(svp64_rm
.spr
)
590 # fiinally yield the svp64 prefix and the thingy. v3.0b opcode
591 rc
= '.' if rc_mode
else ''
592 yield ".long 0x%x" % svp64_prefix
.insn
.value
593 yield "%s %s" % (v30b_op
+rc
, ", ".join(v30b_newfields
))
594 print ("new v3.0B fields", v30b_op
, v30b_newfields
)
596 if __name__
== '__main__':
597 lst
= ['slw 3, 1, 4',
600 'sv.cmpi 5, 1, 3, 2',
602 'sv.isel 64.v, 3, 2, 65.v',
603 'sv.setb/m=r3/sm=1<<r3 5, 31',
604 'sv.setb/vec2 5, 31',
605 'sv.setb/sw=8/ew=16 5, 31',
606 'sv.extsw./ff=eq 5, 31',
607 'sv.extsw./satu/sz/dz/sm=r3/m=r3 5, 31',
608 'sv.extsw./pr=eq 5.v, 31',
609 'sv.add. 5.v, 2.v, 1.v',
615 print ("list", list(isa
))