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 creates
8 an EXT001-encoded "svp64 prefix" (as a .long) 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 Encoding format of arithmetic: https://libre-soc.org/openpower/sv/normal/
15 Encoding format of LDST: https://libre-soc.org/openpower/sv/ldst/
16 **TODO format of branches: https://libre-soc.org/openpower/sv/branches/**
17 **TODO format of CRs: https://libre-soc.org/openpower/sv/cr_ops/**
18 Bugtracker: https://bugs.libre-soc.org/show_bug.cgi?id=578
24 from collections
import OrderedDict
27 from openpower
.decoder
.pseudo
.pagereader
import ISA
28 from openpower
.decoder
.power_svp64
import SVP64RM
, get_regtype
, decode_extra
29 from openpower
.decoder
.selectable_int
import SelectableInt
30 from openpower
.consts
import SVP64MODE
31 from openpower
.decoder
.power_insn
import SVP64Instruction
32 from openpower
.decoder
.power_insn
import Database
33 from openpower
.decoder
.power_insn
import Style
34 from openpower
.decoder
.power_insn
import WordInstruction
35 from openpower
.decoder
.power_enums
import find_wiki_dir
38 from openpower
.util
import log
41 # decode GPR into sv extra
42 def get_extra_gpr(etype
, regmode
, field
):
43 if regmode
== 'scalar':
44 # cut into 2-bits 5-bits SS FFFFF
46 field
= field
& 0b11111
48 # cut into 5-bits 2-bits FFFFF SS
49 sv_extra
= field
& 0b11
51 return sv_extra
, field
54 # decode 3-bit CR into sv extra
55 def get_extra_cr_3bit(etype
, regmode
, field
):
56 if regmode
== 'scalar':
57 # cut into 2-bits 3-bits SS FFF
61 # cut into 3-bits 4-bits FFF SSSS but will cut 2 zeros off later
62 sv_extra
= field
& 0b1111
64 return sv_extra
, field
68 def decode_subvl(encoding
):
69 pmap
= {'2': 0b01, '3': 0b10, '4': 0b11}
70 assert encoding
in pmap
, \
71 "encoding %s for SUBVL not recognised" % encoding
76 def decode_elwidth(encoding
):
77 pmap
= {'8': 0b11, '16': 0b10, '32': 0b01}
78 assert encoding
in pmap
, \
79 "encoding %s for elwidth not recognised" % encoding
83 # decodes predicate register encoding
84 def decode_predicate(encoding
):
95 'nl': (1, 0b001), 'ge': (1, 0b001), # same value
97 'ng': (1, 0b011), 'le': (1, 0b011), # same value
100 'so': (1, 0b110), 'un': (1, 0b110), # same value
101 'ns': (1, 0b111), 'nu': (1, 0b111), # same value
103 assert encoding
in pmap
, \
104 "encoding %s for predicate not recognised" % encoding
105 return pmap
[encoding
]
108 # decodes "Mode" in similar way to BO field (supposed to, anyway)
109 def decode_bo(encoding
):
110 pmap
= { # TODO: double-check that these are the same as Branch BO
112 'nl': 0b001, 'ge': 0b001, # same value
114 'ng': 0b011, 'le': 0b011, # same value
117 'so': 0b110, 'un': 0b110, # same value
118 'ns': 0b111, 'nu': 0b111, # same value
120 assert encoding
in pmap
, \
121 "encoding %s for BO Mode not recognised" % encoding
122 # barse-ackwards MSB0/LSB0. sigh. this would be nice to be the
123 # same as the decode_predicate() CRfield table above, but (inv,CRbit)
124 # is how it is in the spec [decode_predicate is (CRbit,inv)]
125 mapped
= pmap
[encoding
]
126 si
= SelectableInt(0, 3)
127 si
[0] = mapped
& 1 # inv
128 si
[1:3] = mapped
>> 1 # CR
132 # partial-decode fail-first mode
133 def decode_ffirst(encoding
):
134 if encoding
in ['RC1', '~RC1']:
136 return decode_bo(encoding
)
139 def decode_reg(field
, macros
=None):
142 # decode the field number. "5.v" or "3.s" or "9"
143 # and now also "*0", and "*%0". note: *NOT* to add "*%rNNN" etc.
144 # https://bugs.libre-soc.org/show_bug.cgi?id=884#c0
145 if field
.startswith(("*%", "*")):
146 if field
.startswith("*%"):
150 while field
in macros
:
151 field
= macros
[field
]
152 return int(field
), "vector" # actual register number
154 # try old convention (to be retired)
155 field
= field
.split(".")
156 regmode
= 'scalar' # default
160 elif field
[1] == 'v':
162 field
= int(field
[0]) # actual register number
163 return field
, regmode
166 def decode_imm(field
):
167 ldst_imm
= "(" in field
and field
[-1] == ')'
169 return field
[:-1].split("(")
174 def crf_extra(etype
, rname
, extra_idx
, regmode
, field
, extras
):
175 """takes a CR Field number (CR0-CR127), splits into EXTRA2/3 and v3.0
176 the scalar/vector mode (crNN.v or crNN.s) changes both the format
177 of the EXTRA2/3 encoding as well as what range of registers is possible.
178 this function can be used for both BF/BFA and BA/BB/BT by first removing
179 the bottom 2 bits of BA/BB/BT then re-instating them after encoding.
180 see https://libre-soc.org/openpower/sv/svp64/appendix/#cr_extra
183 sv_extra
, field
= get_extra_cr_3bit(etype
, regmode
, field
)
184 # now sanity-check (and shrink afterwards)
185 if etype
== 'EXTRA2':
186 # 3-bit CR Field (BF, BFA) EXTRA2 encoding
187 if regmode
== 'scalar':
188 # range is CR0-CR15 in increments of 1
189 assert (sv_extra
>> 1) == 0, \
190 "scalar CR %s cannot fit into EXTRA2 %s" % \
191 (rname
, str(extras
[extra_idx
]))
192 # all good: encode as scalar
193 sv_extra
= sv_extra
& 0b01
195 # range is CR0-CR127 in increments of 16
196 assert sv_extra
& 0b111 == 0, \
197 "vector CR %s cannot fit into EXTRA2 %s" % \
198 (rname
, str(extras
[extra_idx
]))
199 # all good: encode as vector (bit 2 set)
200 sv_extra
= 0b10 |
(sv_extra
>> 3)
202 # 3-bit CR Field (BF, BFA) EXTRA3 encoding
203 if regmode
== 'scalar':
204 # range is CR0-CR31 in increments of 1
205 assert (sv_extra
>> 2) == 0, \
206 "scalar CR %s cannot fit into EXTRA3 %s" % \
207 (rname
, str(extras
[extra_idx
]))
208 # all good: encode as scalar
209 sv_extra
= sv_extra
& 0b11
211 # range is CR0-CR127 in increments of 8
212 assert sv_extra
& 0b11 == 0, \
213 "vector CR %s cannot fit into EXTRA3 %s" % \
214 (rname
, str(extras
[extra_idx
]))
215 # all good: encode as vector (bit 3 set)
216 sv_extra
= 0b100 |
(sv_extra
>> 2)
217 return sv_extra
, field
220 def to_number(field
):
221 if field
.startswith("0x"):
223 if field
.startswith("0b"):
228 DB
= Database(find_wiki_dir())
231 # decodes svp64 assembly listings and creates EXT001 svp64 prefixes
233 def __init__(self
, lst
, bigendian
=False, macros
=None):
238 self
.trans
= self
.translate(lst
)
239 self
.isa
= ISA() # reads the v3.0B pseudo-code markdown files
240 self
.svp64
= SVP64RM() # reads the svp64 Remap entries for registers
241 assert bigendian
== False, "error, bigendian not supported yet"
244 yield from self
.trans
246 def translate_one(self
, insn
, macros
=None):
249 macros
.update(self
.macros
)
252 insn_no_comments
= insn
.partition('#')[0].strip()
253 if not insn_no_comments
:
256 # find first space, to get opcode
257 ls
= insn_no_comments
.split()
259 # now find opcode fields
260 fields
= ''.join(ls
[1:]).split(',')
261 mfields
= list(filter(bool, map(str.strip
, fields
)))
262 log("opcode, fields", ls
, opcode
, mfields
)
265 for field
in mfields
:
266 fields
.append(macro_subst(macros
, field
))
267 log("opcode, fields substed", ls
, opcode
, fields
)
269 # identify if it is a word instruction
271 if os
.environ
.get("INSNDB"):
273 if record
is not None:
274 insn
= WordInstruction
.assemble(db
=DB
,
275 entry
=opcode
, arguments
=fields
)
277 f
".long 0x{int(insn):08X}",
284 # identify if is a svp64 mnemonic
285 if not opcode
.startswith('sv.'):
286 yield insn
# unaltered
288 opcode
= opcode
[3:] # strip leading "sv"
290 # start working on decoding the svp64 op: sv.basev30Bop/vec2/mode
291 opmodes
= opcode
.split("/") # split at "/"
292 v30b_op_orig
= opmodes
.pop(0) # first is the v3.0B
293 # check instruction ends with dot
294 rc_mode
= v30b_op_orig
.endswith('.')
296 v30b_op
= v30b_op_orig
[:-1]
298 v30b_op
= v30b_op_orig
301 if os
.environ
.get("INSNDB"):
303 if record
is not None:
304 insn
= SVP64Instruction
.assemble(db
=DB
,
308 prefix
= int(insn
.prefix
)
309 suffix
= int(insn
.suffix
)
310 yield f
".long 0x{prefix:08X}"
311 yield from insn
.suffix
.disassemble(db
=DB
, style
=Style
.SHORT
)
314 # look up the 32-bit op (original, with "." if it has it)
315 if v30b_op_orig
in isa
.instr
:
316 isa_instr
= isa
.instr
[v30b_op_orig
]
318 raise Exception("opcode %s of '%s' not supported" %
319 (v30b_op_orig
, insn
))
321 # look up the svp64 op, first the original (with "." if it has it)
322 if v30b_op_orig
in svp64
.instrs
:
323 rm
= svp64
.instrs
[v30b_op_orig
] # one row of the svp64 RM CSV
324 # then without the "." (if there was one)
325 elif v30b_op
in svp64
.instrs
:
326 rm
= svp64
.instrs
[v30b_op
] # one row of the svp64 RM CSV
328 raise Exception(f
"opcode {v30b_op_orig!r} of "
329 f
"{insn!r} not an svp64 instruction")
331 # get regs info e.g. "RT,RA,RB"
332 v30b_regs
= isa_instr
.regs
[0]
333 log("v3.0B op", v30b_op
, "Rc=1" if rc_mode
else '')
334 log("v3.0B regs", opcode
, v30b_regs
)
337 # right. the first thing to do is identify the ordering of
338 # the registers, by name. the EXTRA2/3 ordering is in
339 # rm['0']..rm['3'] but those fields contain the names RA, BB
340 # etc. we have to read the pseudocode to understand which
341 # reg is which in our instruction. sigh.
343 # first turn the svp64 rm into a "by name" dict, recording
344 # which position in the RM EXTRA it goes into
345 # also: record if the src or dest was a CR, for sanity-checking
346 # (elwidth overrides on CRs are banned)
347 decode
= decode_extra(rm
)
348 dest_reg_cr
, src_reg_cr
, svp64_src
, svp64_dest
= decode
350 log("EXTRA field index, src", svp64_src
)
351 log("EXTRA field index, dest", svp64_dest
)
353 # okaaay now we identify the field value (opcode N,N,N) with
354 # the pseudo-code info (opcode RT, RA, RB)
355 assert len(fields
) == len(v30b_regs
), \
356 "length of fields %s must match insn `%s` fields %s" % \
357 (str(v30b_regs
), insn
, str(fields
))
358 opregfields
= zip(fields
, v30b_regs
) # err that was easy
360 # now for each of those find its place in the EXTRA encoding
361 # note there is the possibility (for LD/ST-with-update) of
362 # RA occurring **TWICE**. to avoid it getting added to the
363 # v3.0B suffix twice, we spot it as a duplicate, here
364 extras
= OrderedDict()
365 for idx
, (field
, regname
) in enumerate(opregfields
):
366 imm
, regname
= decode_imm(regname
)
367 rtype
= get_regtype(regname
)
368 log(" idx find", rtype
, idx
, field
, regname
, imm
)
370 # probably an immediate field, append it straight
371 extras
[('imm', idx
, False)] = (idx
, field
, None, None, None)
373 extra
= svp64_src
.get(regname
, None)
374 if extra
is not None:
375 extra
= ('s', extra
, False) # not a duplicate
376 extras
[extra
] = (idx
, field
, regname
, rtype
, imm
)
377 log(" idx src", idx
, extra
, extras
[extra
])
378 dextra
= svp64_dest
.get(regname
, None)
379 log("regname in", regname
, dextra
)
380 if dextra
is not None:
381 is_a_duplicate
= extra
is not None # duplicate spotted
382 dextra
= ('d', dextra
, is_a_duplicate
)
383 extras
[dextra
] = (idx
, field
, regname
, rtype
, imm
)
384 log(" idx dst", idx
, extra
, extras
[dextra
])
386 # great! got the extra fields in their associated positions:
387 # also we know the register type. now to create the EXTRA encodings
388 etype
= rm
['Etype'] # Extra type: EXTRA3/EXTRA2
389 ptype
= rm
['Ptype'] # Predication type: Twin / Single
392 for extra_idx
, (idx
, field
, rname
, rtype
, iname
) in extras
.items():
393 # is it a field we don't alter/examine? if so just put it
396 v30b_newfields
.append(field
)
399 # identify if this is a ld/st immediate(reg) thing
400 ldst_imm
= "(" in field
and field
[-1] == ')'
402 immed
, field
= field
[:-1].split("(")
404 field
, regmode
= decode_reg(field
, macros
=macros
)
405 log(" ", extra_idx
, rname
, rtype
,
406 regmode
, iname
, field
, end
=" ")
408 # see Mode field https://libre-soc.org/openpower/sv/svp64/
409 # XXX TODO: the following is a bit of a laborious repeated
410 # mess, which could (and should) easily be parameterised.
411 # XXX also TODO: the LD/ST modes which are different
412 # https://libre-soc.org/openpower/sv/ldst/
414 # rright. SVP64 register numbering is from 0 to 127
415 # for GPRs, FPRs *and* CR Fields, where for v3.0 the GPRs and RPFs
416 # are 0-31 and CR Fields are only 0-7. the SVP64 RM "Extra"
417 # area is used to extend the numbering from the 32-bit
418 # instruction, and also to record whether the register
419 # is scalar or vector. on a per-operand basis. this
420 # results in a slightly finnicky encoding: here we go...
422 # encode SV-GPR and SV-FPR field into extra, v3.0field
423 if rtype
in ['GPR', 'FPR']:
424 sv_extra
, field
= get_extra_gpr(etype
, regmode
, field
)
425 # now sanity-check. EXTRA3 is ok, EXTRA2 has limits
426 # (and shrink to a single bit if ok)
427 if etype
== 'EXTRA2':
428 if regmode
== 'scalar':
429 # range is r0-r63 in increments of 1
430 assert (sv_extra
>> 1) == 0, \
431 "scalar GPR %s cannot fit into EXTRA2 %s" % \
432 (rname
, str(extras
[extra_idx
]))
433 # all good: encode as scalar
434 sv_extra
= sv_extra
& 0b01
436 # range is r0-r127 in increments of 2 (r0 r2 ... r126)
437 assert sv_extra
& 0b01 == 0, \
438 "%s: vector field %s cannot fit " \
440 (insn
, rname
, str(extras
[extra_idx
]))
441 # all good: encode as vector (bit 2 set)
442 sv_extra
= 0b10 |
(sv_extra
>> 1)
443 elif regmode
== 'vector':
444 # EXTRA3 vector bit needs marking
447 # encode SV-CR 3-bit field into extra, v3.0field.
448 # 3-bit is for things like BF and BFA
449 elif rtype
== 'CR_3bit':
450 sv_extra
, field
= crf_extra(etype
, rname
, extra_idx
,
451 regmode
, field
, extras
)
453 # encode SV-CR 5-bit field into extra, v3.0field
454 # 5-bit is for things like BA BB BC BT etc.
455 # *sigh* this is the same as 3-bit except the 2 LSBs of the
456 # 5-bit field are passed through unaltered.
457 elif rtype
== 'CR_5bit':
458 cr_subfield
= field
& 0b11 # record bottom 2 bits for later
459 field
= field
>> 2 # strip bottom 2 bits
460 # use the exact same 3-bit function for the top 3 bits
461 sv_extra
, field
= crf_extra(etype
, rname
, extra_idx
,
462 regmode
, field
, extras
)
463 # reconstruct the actual 5-bit CR field (preserving the
464 # bottom 2 bits, unaltered)
465 field
= (field
<< 2) | cr_subfield
468 raise Exception("no type match: %s" % rtype
)
470 # capture the extra field info
471 log("=>", "%5s" % bin(sv_extra
), field
)
472 extras
[extra_idx
] = sv_extra
474 # append altered field value to v3.0b, differs for LDST
475 # note that duplicates are skipped e.g. EXTRA2 contains
476 # *BOTH* s:RA *AND* d:RA which happens on LD/ST-with-update
477 srcdest
, idx
, duplicate
= extra_idx
478 if duplicate
: # skip adding to v3.0b fields, already added
481 v30b_newfields
.append(("%s(%s)" % (immed
, str(field
))))
483 v30b_newfields
.append(str(field
))
485 log("new v3.0B fields", v30b_op
, v30b_newfields
)
486 log("extras", extras
)
488 # rright. now we have all the info. start creating SVP64 instruction.
489 svp64_insn
= SVP64Instruction
.pair(prefix
=0, suffix
=0)
490 svp64_prefix
= svp64_insn
.prefix
491 svp64_rm
= svp64_insn
.prefix
.rm
493 # begin with EXTRA fields
494 for idx
, sv_extra
in extras
.items():
500 srcdest
, idx
, duplicate
= idx
501 if etype
== 'EXTRA2':
502 svp64_rm
.extra2
[idx
] = sv_extra
504 svp64_rm
.extra3
[idx
] = sv_extra
506 # identify if the op is a LD/ST.
507 # see https://libre-soc.org/openpower/sv/ldst/
508 is_ldst
= rm
['mode'] in ['LDST_IDX', 'LDST_IMM']
509 is_ldst_idx
= rm
['mode'] == 'LDST_IDX'
510 is_ldst_imm
= rm
['mode'] == 'LDST_IMM'
511 is_ld
= v30b_op
.startswith("l") and is_ldst
512 is_st
= v30b_op
.startswith("s") and is_ldst
514 # branch-conditional or CR detection
515 is_bc
= rm
['mode'] == 'BRANCH'
516 is_cr
= rm
['mode'] == 'CROP'
521 destwid
= 0 # bits 4-5
522 srcwid
= 0 # bits 6-7
524 smask
= 0 # bits 16-18 but only for twin-predication
525 mode
= 0 # bits 19-23
527 mask_m_specified
= False
538 mapreduce_crm
= False
549 # ok let's start identifying opcode augmentation fields
550 for encmode
in opmodes
:
551 # predicate mask (src and dest)
552 if encmode
.startswith("m="):
554 pmmode
, pmask
= decode_predicate(encmode
[2:])
555 smmode
, smask
= pmmode
, pmask
557 mask_m_specified
= True
558 # predicate mask (dest)
559 elif encmode
.startswith("dm="):
561 pmmode
, pmask
= decode_predicate(encmode
[3:])
564 # predicate mask (src, twin-pred)
565 elif encmode
.startswith("sm="):
567 smmode
, smask
= decode_predicate(encmode
[3:])
571 elif encmode
.startswith("vec"):
572 subvl
= decode_subvl(encmode
[3:])
573 # elwidth (both src and dest, like mask)
574 elif encmode
.startswith("w="):
575 destwid
= decode_elwidth(encmode
[2:])
576 srcwid
= decode_elwidth(encmode
[2:])
578 elif encmode
.startswith("dw="):
579 destwid
= decode_elwidth(encmode
[3:])
581 elif encmode
.startswith("sw="):
582 srcwid
= decode_elwidth(encmode
[3:])
584 elif encmode
== 'pi':
586 # in indexed mode, set sv_mode=0b00
587 assert is_ldst_imm
is True
589 # element-strided LD/ST
590 elif encmode
== 'els':
592 # in indexed mode, set sv_mode=0b01
596 elif encmode
== 'sats':
597 assert sv_mode
is None
600 elif encmode
== 'satu':
601 assert sv_mode
is None
605 elif encmode
== 'zz': # TODO, a lot more checking on legality
606 dst_zero
= 1 # NOT on cr_ops, that's RM[6]
608 elif encmode
== 'sz':
610 elif encmode
== 'dz':
613 elif encmode
.startswith("ff="):
614 assert sv_mode
is None
615 if is_cr
: # sigh, CROPs is different
619 failfirst
= decode_ffirst(encmode
[3:])
620 assert sea
is False, "cannot use failfirst with signed-address"
621 # predicate-result, interestingly same as fail-first
622 elif encmode
.startswith("pr="):
623 assert sv_mode
is None
625 predresult
= decode_ffirst(encmode
[3:])
626 # map-reduce mode, reverse-gear
627 elif encmode
== 'mrr':
628 assert sv_mode
is None
633 elif encmode
== 'mr':
634 assert sv_mode
is None
637 elif encmode
== 'crm': # CR on map-reduce
638 assert sv_mode
is None
641 elif encmode
== 'vli':
642 assert failfirst
is not False, "VLi only allowed in failfirst"
644 elif encmode
== 'sea':
647 assert failfirst
is False, "cannot use ffirst+signed-address"
650 svp64_rm
.branch
.ALL
= 1
651 elif encmode
== 'snz':
652 svp64_rm
.branch
.sz
= 1
653 svp64_rm
.branch
.SNZ
= 1
654 elif encmode
== 'sl':
655 svp64_rm
.branch
.SL
= 1
656 elif encmode
== 'slu':
657 svp64_rm
.branch
.SLu
= 1
658 elif encmode
== 'lru':
659 svp64_rm
.branch
.LRu
= 1
660 elif encmode
== 'vs':
661 svp64_rm
.branch
.VLS
= 1
662 elif encmode
== 'vsi':
663 svp64_rm
.branch
.VLS
= 1
664 svp64_rm
.branch
.vls
.VLi
= 1
665 elif encmode
== 'vsb':
666 svp64_rm
.branch
.VLS
= 1
667 svp64_rm
.branch
.vls
.VSb
= 1
668 elif encmode
== 'vsbi':
669 svp64_rm
.branch
.VLS
= 1
670 svp64_rm
.branch
.vls
.VSb
= 1
671 svp64_rm
.branch
.vls
.VLi
= 1
672 elif encmode
== 'ctr':
673 svp64_rm
.branch
.CTR
= 1
674 elif encmode
== 'cti':
675 svp64_rm
.branch
.CTR
= 1
676 svp64_rm
.branch
.ctr
.CTi
= 1
678 raise AssertionError("unknown encmode %s" % encmode
)
680 raise AssertionError("unknown encmode %s" % encmode
)
682 # post-inc only available on ld-with-update
684 assert "u" in opcode
, "/pi only available on ld/st-update"
686 # sanity check if dz/zz used in branch-mode
687 if is_bc
and dst_zero
:
688 raise AssertionError("dz/zz not supported in branch, use 'sz'")
690 # check sea *after* all qualifiers are evaluated
692 assert sv_mode
in (None, 0b00, 0b01)
695 # since m=xx takes precedence (overrides) sm=xx and dm=xx,
696 # treat them as mutually exclusive
698 assert not has_smask
,\
699 "cannot have both source-mask and predicate mask"
700 assert not has_pmask
,\
701 "cannot have both dest-mask and predicate mask"
702 # since the default is INT predication (ALWAYS), if you
703 # specify one CR mask, you must specify both, to avoid
704 # mixing INT and CR reg types
705 if has_pmask
and pmmode
== 1:
707 "need explicit source-mask in CR twin predication"
708 if has_smask
and smmode
== 1:
710 "need explicit dest-mask in CR twin predication"
711 # sanity-check that 2Pred mask is same mode
712 if has_pmask
and has_smask
:
713 assert smmode
== pmmode
, \
714 "predicate masks %s and %s must be same reg type" % \
717 # sanity-check that twin-predication mask only specified in 2P mode
719 assert not has_smask
, \
720 "source-mask can only be specified on Twin-predicate ops"
721 assert not has_pmask
, \
722 "dest-mask can only be specified on Twin-predicate ops"
724 # construct the mode field, doing sanity-checking along the way
726 assert has_smask
or mask_m_specified
, \
727 "src zeroing requires a source predicate"
729 assert has_pmask
or mask_m_specified
, \
730 "dest zeroing requires a dest predicate"
732 # okaaay, so there are 4 different modes, here, which will be
733 # partly-merged-in: is_ldst is merged in with "normal", but
734 # is_bc is so different it's done separately. likewise is_cr
735 # (when it is done). here are the maps:
737 # for "normal" arithmetic: https://libre-soc.org/openpower/sv/normal/
739 | 0-1 | 2 | 3 4 | description |
740 | --- | --- |---------|-------------------------- |
741 | 00 | 0 | dz sz | simple mode |
742 | 00 | 1 | 0 RG | scalar reduce mode (mapreduce) |
743 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
744 | 01 | inv | VLi RC1 | Rc=0: ffirst z/nonz |
745 | 10 | N | dz sz | sat mode: N=0/1 u/s |
746 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
747 | 11 | inv | zz RC1 | Rc=0: pred-result z/nonz |
750 # https://libre-soc.org/openpower/sv/ldst/
751 # for LD/ST-immediate:
753 | 0-1 | 2 | 3 4 | description |
754 | --- | --- |---------|--------------------------- |
755 | 00 | 0 | zz els | normal mode |
756 | 00 | 1 | pi lf | post-inc, LD-fault-first |
757 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
758 | 01 | inv | els RC1 | Rc=0: ffirst z/nonz |
759 | 10 | N | zz els | sat mode: N=0/1 u/s |
760 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
761 | 11 | inv | els RC1 | Rc=0: pred-result z/nonz |
764 # for LD/ST-indexed (RA+RB):
766 | 0-1 | 2 | 3 4 | description |
767 | --- | --- |---------|----------------------------- |
768 | 00 | SEA | dz sz | normal mode |
769 | 01 | SEA | dz sz | strided (scalar only source) |
770 | 10 | N | dz sz | sat mode: N=0/1 u/s |
771 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
772 | 11 | inv | dz RC1 | Rc=0: pred-result z/nonz |
775 # and leaving out branches and cr_ops for now because they're
777 """ TODO branches and cr_ops
781 sv_mode
= int(svp64_rm
.mode
[0, 1])
783 svp64_rm
.branch
.sz
= 1
786 ######################################
787 # "element-strided" mode, ldst_idx
788 if sv_mode
== 0b01 and is_ldst_idx
:
789 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
790 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
791 mode |
= sea
<< SVP64MODE
.SEA
# el-strided
793 ######################################
795 elif sv_mode
is None:
796 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
797 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
799 # TODO: for now, LD/ST-indexed is ignored.
800 mode |
= ldst_elstride
<< SVP64MODE
.ELS_NORMAL
# el-strided
802 # TODO, reduce and subvector mode
803 # 00 1 dz CRM reduce mode (mapreduce), SUBVL=1
804 # 00 1 SVM CRM subvector reduce mode, SUBVL>1
808 ######################################
809 # ldst-immediate "post" (and "load-fault-first" modes)
810 elif sv_mode
== 0b00 and ldst_postinc
== 1: # (or ldst_ld_ffirst)
811 mode |
= (0b1 << SVP64MODE
.LDI_POST
) # sets bit 2
812 mode |
= (ldst_postinc
<< SVP64MODE
.LDI_PI
) # sets post-inc
814 ######################################
816 elif sv_mode
== 0b00:
817 mode |
= (0b1 << SVP64MODE
.REDUCE
) # sets mapreduce
818 assert dst_zero
== 0, "dest-zero not allowed in mapreduce mode"
820 mode |
= (0b1 << SVP64MODE
.RG
) # sets Reverse-gear mode
822 mode |
= (0b1 << SVP64MODE
.CRM
) # sets CRM mode
823 assert rc_mode
, "CRM only allowed when Rc=1"
824 # bit of weird encoding to jam zero-pred or SVM mode in.
825 # SVM mode can be enabled only when SUBVL=2/3/4 (vec2/3/4)
827 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
829 ######################################
831 elif failfirst
is not False and not is_cr
: # sv_mode == 0b01:
832 assert src_zero
== 0, "dest-zero not allowed in failfirst mode"
833 if failfirst
== 'RC1':
834 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
835 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
836 assert rc_mode
== False, "ffirst RC1 only ok when Rc=0"
837 elif failfirst
== '~RC1':
838 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
839 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
840 mode |
= (0b1 << SVP64MODE
.INV
) # ... with inversion
841 assert rc_mode
== False, "ffirst RC1 only ok when Rc=0"
843 assert dst_zero
== 0, "dst-zero not allowed in ffirst BO"
844 assert rc_mode
, "ffirst BO only possible when Rc=1"
845 mode |
= (failfirst
<< SVP64MODE
.BO_LSB
) # set BO
847 # (crops is really different)
848 elif failfirst
is not False and is_cr
:
849 if failfirst
in ['RC1', '~RC1']:
850 mode |
= (src_zero
<< SVP64MODE
.SZ
) # predicate src-zeroing
851 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
852 if failfirst
== '~RC1':
853 mode |
= (0b1 << SVP64MODE
.INV
) # ... with inversion
855 assert dst_zero
== src_zero
, "dz must equal sz in ffirst BO"
856 mode |
= (failfirst
<< SVP64MODE
.BO_LSB
) # set BO
857 svp64_rm
.cr_op
.zz
= dst_zero
859 sv_mode |
= 1 # set VLI in LSB of 2-bit mode
860 #svp64_rm.cr_op.vli = 1
862 ######################################
864 elif sv_mode
== 0b10:
865 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
866 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
867 mode |
= (saturation
<< SVP64MODE
.N
) # signed/us saturation
869 ######################################
870 # "predicate-result" modes. err... code-duplication from ffirst
871 elif sv_mode
== 0b11:
872 assert src_zero
== 0, "dest-zero not allowed in predresult mode"
873 if predresult
== 'RC1':
874 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
875 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
876 assert rc_mode
== False, "pr-mode RC1 only ok when Rc=0"
877 elif predresult
== '~RC1':
878 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
879 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
880 mode |
= (0b1 << SVP64MODE
.INV
) # ... with inversion
881 assert rc_mode
== False, "pr-mode RC1 only ok when Rc=0"
883 assert dst_zero
== 0, "dst-zero not allowed in pr-mode BO"
884 assert rc_mode
, "pr-mode BO only possible when Rc=1"
885 mode |
= (predresult
<< SVP64MODE
.BO_LSB
) # set BO
887 # whewww.... modes all done :)
888 # now put into svp64_rm, but respect MSB0 order
890 mode |
= (0b1 << SVP64MODE
.MOD2_LSB
)
892 mode |
= (0b1 << SVP64MODE
.MOD2_MSB
)
895 mode |
= (0b1 << SVP64MODE
.SEA
)
897 # this is a mess. really look forward to replacing it with Insn DB
899 svp64_rm
.mode
= mode
# mode: bits 19-23
900 if vli
and not is_cr
:
901 svp64_rm
.normal
.ffrc0
.VLi
= 1
903 # put in predicate masks into svp64_rm
905 svp64_rm
.smask
= smask
# source pred: bits 16-18
907 # put in elwidths unless cr
909 svp64_rm
.ewsrc
= srcwid
# srcwid: bits 6-7
910 svp64_rm
.elwidth
= destwid
# destwid: bits 4-5
912 svp64_rm
.mmode
= mmode
# mask mode: bit 0
913 svp64_rm
.mask
= pmask
# 1-pred: bits 1-3
914 svp64_rm
.subvl
= subvl
# and subvl: bits 8-9
916 # nice debug printout. (and now for something completely different)
917 # https://youtu.be/u0WOIwlXE9g?t=146
918 svp64_rm_value
= int(svp64_rm
)
919 log("svp64_rm", hex(svp64_rm_value
), bin(svp64_rm_value
))
920 log(" mmode 0 :", bin(mmode
))
921 log(" pmask 1-3 :", bin(pmask
))
922 log(" dstwid 4-5 :", bin(destwid
))
923 log(" srcwid 6-7 :", bin(srcwid
))
924 log(" subvl 8-9 :", bin(subvl
))
925 log(" mode 19-23:", bin(svp64_rm
.mode
))
926 offs
= 2 if etype
== 'EXTRA2' else 3 # 2 or 3 bits
927 for idx
, sv_extra
in extras
.items():
932 srcdest
, idx
, duplicate
= idx
933 start
= (10+idx
*offs
)
935 log(" extra%d %2d-%2d:" % (idx
, start
, end
),
938 log(" smask 16-17:", bin(smask
))
941 # update prefix PO and ID (aka PID)
942 svp64_prefix
.PO
= 0x1
943 svp64_prefix
.id = 0b11
945 # fiinally yield the svp64 prefix and the thingy. v3.0b opcode
946 rc
= '.' if rc_mode
else ''
947 yield ".long 0x%08x" % int(svp64_prefix
)
948 log(v30b_op
, v30b_newfields
)
951 if not v30b_op
.endswith('.'):
955 if os
.environ
.get("INSNDB"):
957 if record
is not None:
958 insn
= WordInstruction
.assemble(db
=DB
,
959 entry
=opcode
, arguments
=fields
)
961 f
".long 0x{int(insn):08X}",
967 if not v30b_op
.endswith('.'):
969 yield "%s %s" % (v30b_op
, ", ".join(v30b_newfields
))
970 for (name
, span
) in svp64_insn
.traverse("SVP64"):
971 value
= svp64_insn
.storage
[span
]
972 log(name
, f
"{value.value:0{value.bits}b}", span
)
973 log("new v3.0B fields", v30b_op
, v30b_newfields
)
975 def translate(self
, lst
):
977 yield from self
.translate_one(insn
)
980 def macro_subst(macros
, txt
):
982 log("subst", txt
, macros
)
985 for macro
, value
in macros
.items():
988 replaced
= txt
.replace(macro
, value
)
989 log("macro", txt
, "replaced", replaced
, macro
, value
)
992 toreplace
= '%s.s' % macro
995 replaced
= txt
.replace(toreplace
, "%s.s" % value
)
996 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
999 toreplace
= '%s.v' % macro
1000 if toreplace
== txt
:
1002 replaced
= txt
.replace(toreplace
, "%s.v" % value
)
1003 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1006 toreplace
= '*%s' % macro
1007 if toreplace
in txt
:
1009 replaced
= txt
.replace(toreplace
, '*%s' % value
)
1010 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1013 toreplace
= '(%s)' % macro
1014 if toreplace
in txt
:
1016 replaced
= txt
.replace(toreplace
, '(%s)' % value
)
1017 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1020 log(" processed", txt
)
1028 if not line
[0].isspace():
1036 # get an input file and an output file
1040 outfile
= sys
.stdout
1041 # read the whole lot in advance in case of in-place
1042 lines
= list(infile
.readlines())
1043 elif len(args
) != 2:
1044 print("pysvp64asm [infile | -] [outfile | -]", file=sys
.stderr
)
1050 infile
= open(args
[0], "r")
1051 # read the whole lot in advance in case of in-place overwrite
1052 lines
= list(infile
.readlines())
1055 outfile
= sys
.stdout
1057 outfile
= open(args
[1], "w")
1059 # read the line, look for custom insn, process it
1060 macros
= {} # macros which start ".set"
1063 op
= line
.split("#")[0].strip()
1065 if op
.startswith(".set"):
1066 macro
= op
[4:].split(",")
1067 (macro
, value
) = map(str.strip
, macro
)
1068 macros
[macro
] = value
1070 if not op
or op
.startswith("#"):
1073 (ws
, line
) = get_ws(line
)
1074 lst
= isa
.translate_one(op
, macros
)
1075 lst
= '; '.join(lst
)
1076 outfile
.write("%s%s # %s\n" % (ws
, lst
, op
))
1079 if __name__
== '__main__':
1080 lst
= ['slw 3, 1, 4',
1083 'sv.cmpi 5, 1, 3, 2',
1085 'sv.isel 64.v, 3, 2, 65.v',
1086 'sv.setb/dm=r3/sm=1<<r3 5, 31',
1087 'sv.setb/m=r3 5, 31',
1088 'sv.setb/vec2 5, 31',
1089 'sv.setb/sw=8/ew=16 5, 31',
1090 'sv.extsw./ff=eq 5, 31',
1091 'sv.extsw./satu/sz/dz/sm=r3/dm=r3 5, 31',
1092 'sv.extsw./pr=eq 5.v, 31',
1093 'sv.add. 5.v, 2.v, 1.v',
1094 'sv.add./m=r3 5.v, 2.v, 1.v',
1097 'sv.stw 5.v, 4(1.v)',
1098 'sv.ld 5.v, 4(1.v)',
1099 'setvl. 2, 3, 4, 0, 1, 1',
1100 'sv.setvl. 2, 3, 4, 0, 1, 1',
1103 "sv.stfsu 0.v, 16(4.v)",
1106 "sv.stfsu/els 0.v, 16(4)",
1109 'sv.add./mr 5.v, 2.v, 1.v',
1111 macros
= {'win2': '50', 'win': '60'}
1113 'sv.addi win2.v, win.v, -1',
1114 'sv.add./mrr 5.v, 2.v, 1.v',
1115 #'sv.lhzsh 5.v, 11(9.v), 15',
1116 #'sv.lwzsh 5.v, 11(9.v), 15',
1117 'sv.ffmadds 6.v, 2.v, 4.v, 6.v',
1120 #'sv.fmadds 0.v, 8.v, 16.v, 4.v',
1121 #'sv.ffadds 0.v, 8.v, 4.v',
1122 'svremap 11, 0, 1, 2, 3, 2, 1',
1123 'svshape 8, 1, 1, 1, 0',
1124 'svshape 8, 1, 1, 1, 1',
1127 #'sv.lfssh 4.v, 11(8.v), 15',
1128 #'sv.lwzsh 4.v, 11(8.v), 15',
1129 #'sv.svstep. 2.v, 4, 0',
1130 #'sv.fcfids. 48.v, 64.v',
1131 'sv.fcoss. 80.v, 0.v',
1132 'sv.fcoss. 20.v, 0.v',
1135 'sv.bc/all 3,12,192',
1136 'sv.bclr/vsbi 3,81.v,192',
1137 'sv.ld 5.v, 4(1.v)',
1138 'sv.svstep. 2.v, 4, 0',
1149 'svindex 0,0,1,0,0,0,0',
1152 'sv.svstep./m=r3 2.v, 4, 0',
1153 'ternlogi 0,0,0,0x5',
1165 'sv.andi. *80, *80, 1',
1166 'sv.ffmadds. 6.v, 2.v, 4.v, 6.v', # incorrectly inserted 32-bit op
1167 'sv.ffmadds 6.v, 2.v, 4.v, 6.v', # correctly converted to .long
1168 'svshape2 8, 1, 31, 7, 1, 1',
1169 'sv.ld 5.v, 4(1.v)',
1170 'sv.stw 5.v, 4(1.v)',
1171 'sv.bc/all 3,12,192',
1175 #"sv.cmp/ff=gt *0,*1,*2,0",
1179 isa
= SVP64Asm(lst
, macros
=macros
)
1180 log("list:\n", "\n\t".join(list(isa
)))
1181 # running svp64.py is designed to test hard-coded lists
1182 # (above) - which strictly speaking should all be unit tests.
1183 # if you need to actually do assembler translation at the
1184 # commandline use "pysvp64asm" - see setup.py
1185 # XXX NO. asm_process()