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_enums
import find_wiki_dir
36 from openpower
.util
import log
39 def instruction(*fields
):
40 def instruction(insn
, desc
):
41 (value
, start
, end
) = desc
42 bits
= ((1,) * ((end
+ 1) - start
))
45 mask
= ((mask
<< 1) | bit
)
46 return (insn |
((value
& mask
) << (31 - end
)))
48 return functools
.reduce(instruction
, fields
, 0)
54 def _insn(name
, **kwargs
):
58 def _custom_insns(*insns
):
59 """ a decorator that adds the function to `CUSTOM_INSNS` """
64 insns_
= (fn
.__name
__, {}),
67 for name
, kwargs
in insns_
:
68 if not isinstance(name
, str):
69 raise TypeError("instruction name must be a str: {name!r}")
70 if name
in CUSTOM_INSNS
:
71 raise ValueError(f
"duplicate instruction mnemonic: {name!r}")
72 # use getcallargs to check that arguments work:
73 inspect
.getcallargs(fn
, FIELDS_ARG
, **kwargs
)
74 CUSTOM_INSNS
[name
] = functools
.partial(fn
, **kwargs
)
81 _insn("setvl.", Rc
=1),
83 def setvl(fields
, Rc
):
85 setvl is a *32-bit-only* instruction. It controls SVSTATE.
86 It is *not* a 64-bit-prefixed Vector instruction (no sv.setvl, yet),
87 it is a Vector *control* instruction.
89 * setvl RT,RA,SVi,vf,vs,ms
91 1.6.28 SVL-FORM - from fields.txt
92 |0 |6 |11 |16 |23 |24 |25 |26 |31 |
93 | PO | RT | RA | SVi |ms |vs |vf | XO |Rc |
97 # ARRRGH these are in a non-obvious order in openpower/isa/simplev.mdwn
98 # compared to the SVL-Form above. sigh
99 # setvl RT,RA,SVi,vf,vs,ms
100 (RT
, RA
, SVi
, vf
, vs
, ms
) = fields
116 _insn("svstep", Rc
=0),
117 _insn("svstep.", Rc
=1),
119 def svstep(fields
, Rc
):
121 svstep is a 32-bit instruction. It updates SVSTATE.
122 It *can* be SVP64-prefixed, to indicate that its registers
127 # 1.6.28 SVL-FORM - from fields.txt
128 # |0 |6 |11 |16 |23 |24 |25 |26 |31 |
129 # | PO | RT | / | SVi |/ |/ |vf | XO |Rc |
134 (RT
, SVi
, vf
) = fields
152 svshape is a *32-bit-only* instruction. It updates SVSHAPE and SVSTATE.
153 It is *not* a 64-bit-prefixed Vector instruction (no sv.svshape, yet),
154 it is a Vector *control* instruction.
156 https://libre-soc.org/openpower/sv/remap/#svshape
158 * svshape SVxd,SVyd,SVzd,SVrm,vf
160 # 1.6.33 SVM-FORM from fields.txt
161 # |0 |6 |11 |16 |21 |25 |26 |31 |
162 # | PO | SVxd | SVyd | SVzd | SVrm |vf | XO |
164 note that SVrm is not permitted to be 0b0111, 0b1000 or 0b1001.
165 0b0111 is reserved and 0b100- is for svshape2
170 (SVxd
, SVyd
, SVzd
, SVrm
, vf
) = fields
175 # check SVrm for reserved (and svshape2) values
176 assert SVrm
not in [0b1000, 0b1001], \
177 "svshape reserved SVrm value %s" % bin(SVrm
)
191 def svshape2(fields
):
193 svshape2 is a *32-bit-only* instruction. It updates SVSHAPE and SVSTATE.
194 It is *not* a 64-bit-prefixed Vector instruction (no sv.svshape2, yet),
195 it is a Vector *control* instruction, and is a sort-of hybrid of
196 svshape and svindex, with the key important feature being the "offset".
198 https://libre-soc.org/openpower/sv/remap/discussion
200 * svshape2 SVo,SVM2yx,rmm,SVd,sk,mm
202 # 1.6.35.1 SVM2-FORM from fields.txt
203 # |0 |6 |10 |11 |16 |21 |24|25 |26 |31 |
204 # | PO | SVo |SVMyx| rmm | SVd |XO |mm|sk | XO |
206 note that this fits into the space of svshape and that XO is
207 split across 2 areas.
212 XO2
= 0b100 # not really XO2 but hey
213 (offs
, yx
, rmm
, SVd
, sk
, mm
) = fields
214 SVd
-= 1 # offset by one
218 (offs
, 6, 9), # offset (the whole point of adding svshape2)
219 (yx
, 10, 10), # like svindex
220 (rmm
, 11, 15), # ditto svindex
221 (SVd
, 16, 20), # ditto svindex
222 (XO2
, 21, 23), # actually XO split across 2 places...
223 (mm
, 24, 24), # ditto svindex
224 (sk
, 25, 25), # ditto svindex
232 svindex is a *32-bit-only* instruction. It is a convenience
233 instruction that reduces instruction count for Indexed REMAP
235 It is *not* a 64-bit-prefixed Vector instruction (no sv.svindex, yet),
236 it is a Vector *control* instruction.
239 |0 |6 |11 |16 |21 |23|24|25|26 31|
240 | PO | SVG|rmm | SVd |ew |yx|mm|sk| XO |
242 # note that the dimension field one subtracted
245 (SVG
, rmm
, SVd
, ew
, yx
, mm
, sk
) = fields
263 this is a *32-bit-only* instruction. It updates the SVSHAPE SPR
264 it is *not* a 64-bit-prefixed Vector instruction (no sv.svremap),
265 it is a Vector *control* instruction.
267 * svremap SVme,mi0,mi1,mi2,mo0,mo1,pst
270 |0 |6 |11 |13 |15 |17 |19 |21 |22 |26 |31 |
271 | PO | SVme |mi0 | mi1 | mi2 | mo0 | mo1 |pst |/// | XO |
276 (SVme
, mi0
, mi1
, mi2
, mo0
, mo1
, pst
) = fields
291 # ok from here-on down these are added as 32-bit instructions
292 # and are here only because binutils (at present) doesn't have
293 # them (that's being fixed!)
294 # they can - if implementations then choose - be Vectorised
295 # because they are general-purpose scalar instructions
300 |0 |6 |11 |16 |21 |26 |27 31|
301 | PO | RT | RA | RB |bm |L | XO |
305 (RT
, RA
, RB
, bm
, L
) = fields
317 def _fptrans_insn(name
, XO
):
319 _insn(name
, PO
=63, Rc
=0, XO
=XO
),
320 _insn(name
+ ".", PO
=63, Rc
=1, XO
=XO
),
321 _insn(name
+ "s", PO
=59, Rc
=0, XO
=XO
),
322 _insn(name
+ "s.", PO
=59, Rc
=1, XO
=XO
),
327 *_fptrans_insn("fatan2", XO
=0b1001001110),
328 *_fptrans_insn("fatan2pi", XO
=0b1000001110),
329 *_fptrans_insn("fpow", XO
=0b1111101101),
330 *_fptrans_insn("fpown", XO
=0b1101101100),
331 *_fptrans_insn("fpowr", XO
=0b1111101100),
332 *_fptrans_insn("frootn", XO
=0b1101101101),
333 *_fptrans_insn("fhypot", XO
=0b1010001110),
334 *_fptrans_insn("fminnum08", XO
=0b1011001100),
335 *_fptrans_insn("fmaxnum08", XO
=0b1011101100),
336 *_fptrans_insn("fmin19", XO
=0b1011001101),
337 *_fptrans_insn("fmax19", XO
=0b1011101101),
338 *_fptrans_insn("fminnum19", XO
=0b1011001110),
339 *_fptrans_insn("fmaxnum19", XO
=0b1011101110),
340 *_fptrans_insn("fminc", XO
=0b1011001111),
341 *_fptrans_insn("fmaxc", XO
=0b1011101111),
342 *_fptrans_insn("fminmagnum08", XO
=0b1100001110),
343 *_fptrans_insn("fmaxmagnum08", XO
=0b1100001111),
344 *_fptrans_insn("fminmag19", XO
=0b1101101110),
345 *_fptrans_insn("fmaxmag19", XO
=0b1101101111),
346 *_fptrans_insn("fminmagnum19", XO
=0b1110001110),
347 *_fptrans_insn("fmaxmagnum19", XO
=0b1110001111),
348 *_fptrans_insn("fminmagc", XO
=0b1111101110),
349 *_fptrans_insn("fmaxmagc", XO
=0b1111101111),
350 *_fptrans_insn("fmod", XO
=0b1101001111),
351 *_fptrans_insn("fremainder", XO
=0b1111001111),
353 def fptrans_binary(fields
, PO
, Rc
, XO
):
354 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
355 # however we are out of space with opcode 22
357 # |0 |6 |11 |16 |21 |31 |
358 # | PO | FRT | FRA | FRB | XO |Rc |
359 # | PO | FRT | FRA | RB | XO |Rc |
360 (FRT
, FRA
, FRB
) = fields
372 *_fptrans_insn("frsqrt", XO
=0b1001001100),
373 *_fptrans_insn("fcbrt", XO
=0b1000001100),
374 *_fptrans_insn("frecip", XO
=0b1010001100),
375 *_fptrans_insn("fexp2m1", XO
=0b1100001100),
376 *_fptrans_insn("flog2p1", XO
=0b1100001101),
377 *_fptrans_insn("fexp2", XO
=0b1110001100),
378 *_fptrans_insn("flog2", XO
=0b1110001101),
379 *_fptrans_insn("fexpm1", XO
=0b1100101100),
380 *_fptrans_insn("flogp1", XO
=0b1100101101),
381 *_fptrans_insn("fexp", XO
=0b1110101100),
382 *_fptrans_insn("flog", XO
=0b1110101101),
383 *_fptrans_insn("fexp10m1", XO
=0b1101001100),
384 *_fptrans_insn("flog10p1", XO
=0b1101001101),
385 *_fptrans_insn("fexp10", XO
=0b1111001100),
386 *_fptrans_insn("flog10", XO
=0b1111001101),
387 *_fptrans_insn("fsin", XO
=0b1001001101),
388 *_fptrans_insn("fcos", XO
=0b1001101100),
389 *_fptrans_insn("ftan", XO
=0b1001101101),
390 *_fptrans_insn("fasin", XO
=0b1001001111),
391 *_fptrans_insn("facos", XO
=0b1001101110),
392 *_fptrans_insn("fatan", XO
=0b1001101111),
393 *_fptrans_insn("fsinpi", XO
=0b1000001101),
394 *_fptrans_insn("fcospi", XO
=0b1000101100),
395 *_fptrans_insn("ftanpi", XO
=0b1000101101),
396 *_fptrans_insn("fasinpi", XO
=0b1000001111),
397 *_fptrans_insn("facospi", XO
=0b1000101110),
398 *_fptrans_insn("fatanpi", XO
=0b1000101111),
399 *_fptrans_insn("fsinh", XO
=0b1010001101),
400 *_fptrans_insn("fcosh", XO
=0b1010101100),
401 *_fptrans_insn("ftanh", XO
=0b1010101101),
402 *_fptrans_insn("fasinh", XO
=0b1010001111),
403 *_fptrans_insn("facosh", XO
=0b1010101110),
404 *_fptrans_insn("fatanh", XO
=0b1010101111),
406 def fptrans_unary(fields
, PO
, Rc
, XO
):
407 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
408 # however we are out of space with opcode 22
410 # |0 |6 |11 |16 |21 |31 |
411 # | PO | FRT | /// | FRB | XO |Rc |
424 _insn("ternlogi", Rc
=0),
425 _insn("ternlogi.", Rc
=1),
427 def ternlogi(fields
, Rc
):
428 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
429 # however we are out of space with opcode 22
431 # |0 |6 |11 |16 |21 |29 |31 |
432 # | PO | RT | RA | RB | TLI | XO |Rc |
435 (RT
, RA
, RB
, TLI
) = fields
448 _insn("grev", Rc
=0, imm
=0, word
=0),
449 _insn("grevw", Rc
=0, imm
=0, word
=1),
450 _insn("grevi", Rc
=0, imm
=1, word
=0),
451 _insn("grevwi", Rc
=0, imm
=1, word
=1),
452 _insn("grev.", Rc
=1, imm
=0, word
=0),
453 _insn("grevw.", Rc
=1, imm
=0, word
=1),
454 _insn("grevi.", Rc
=1, imm
=1, word
=0),
455 _insn("grevwi.", Rc
=1, imm
=1, word
=1),
457 def grev(fields
, Rc
, imm
, word
):
458 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
459 # however we are out of space with opcode 22
461 # _ matches fields in table at:
462 # https://libre-soc.org/openpower/sv/bitmanip/
468 (RT
, RA
, XBI
) = fields
469 insn
= (insn
<< 5) | RT
470 insn
= (insn
<< 5) | RA
473 insn
= (insn
<< 6) | XBI
474 insn
= (insn
<< 9) | XO
477 insn
= (insn
<< 5) | XBI
478 insn
= (insn
<< 10) | XO
479 insn
= (insn
<< 1) | Rc
484 _insn("maxs", XO
=0b0111001110, Rc
=0),
485 _insn("maxs.", XO
=0b0111001110, Rc
=1),
486 _insn("maxu", XO
=0b0011001110, Rc
=0),
487 _insn("maxu.", XO
=0b0011001110, Rc
=1),
488 _insn("minu", XO
=0b0001001110, Rc
=0),
489 _insn("minu.", XO
=0b0001001110, Rc
=1),
490 _insn("mins", XO
=0b0101001110, Rc
=0),
491 _insn("mins.", XO
=0b0101001110, Rc
=1),
492 _insn("absdu", XO
=0b1011110110, Rc
=0),
493 _insn("absdu.", XO
=0b1011110110, Rc
=1),
494 _insn("absds", XO
=0b1001110110, Rc
=0),
495 _insn("absds.", XO
=0b1001110110, Rc
=1),
496 _insn("avgadd", XO
=0b1101001110, Rc
=0),
497 _insn("avgadd.", XO
=0b1101001110, Rc
=1),
498 _insn("absdacu", XO
=0b1111110110, Rc
=0),
499 _insn("absdacu.", XO
=0b1111110110, Rc
=1),
500 _insn("absdacs", XO
=0b0111110110, Rc
=0),
501 _insn("absdacs.", XO
=0b0111110110, Rc
=1),
502 _insn("cprop", XO
=0b0110001110, Rc
=0),
503 _insn("cprop.", XO
=0b0110001110, Rc
=1),
505 def av(fields
, XO
, Rc
):
507 # |0 |6 |7|8|9 |10 |11|12|13 |15|16|17 |20|21 |31 |
508 # | PO | RT | RA | RB | XO |Rc |
510 (RT
, RA
, RB
) = fields
523 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
524 # V3.0B 1.6.6 DX-FORM
525 # |0 |6 |7|8|9 |10 |11|12|13 |15|16|17 |26|27 |31 |
526 # | PO | FRS | d1 | d0 | XO |d2 |
530 # first split imm into d1, d0 and d2. sigh
531 d2
= (imm
& 1) # LSB (0)
532 d1
= (imm
>> 1) & 0b11111 # bits 1-5
533 d0
= (imm
>> 6) # MSBs 6-15
546 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
547 # V3.0B 1.6.6 DX-FORM
548 # |0 |6 |7|8|9 |10 |11|12|13 |15|16|17 |26|27 |31 |
549 # | PO | FRS | d1 | d0 | XO |d2 |
553 # first split imm into d1, d0 and d2. sigh
554 d2
= (imm
& 1) # LSB (0)
555 d1
= (imm
>> 1) & 0b11111 # bits 1-5
556 d0
= (imm
>> 6) # MSBs 6-15
568 _insn("pcdec."), # named "pcdec." because it always writes to CR0
571 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
573 # |0 |6 |11 |16 |21 |24|26 |31 |
574 # | PO | RT | RA | RB | RC | XO |once|
577 (RT
, RA
, RB
, RC
, once
) = fields
590 _insn("madded", XO
=50),
591 _insn("divmod2du", XO
=52),
593 def va_form(fields
, XO
):
594 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
596 # |0 |6 |11 |16 |21 |26 |
597 # | PO | RT | RA | RB | RC | XO |
599 (RT
, RA
, RB
, RC
) = fields
611 _insn("dsld", XO
=0b00111001, Rc
=0),
612 _insn("dsld.", XO
=0b00111001, Rc
=1),
613 _insn("dsrd", XO
=0b10111001, Rc
=0),
614 _insn("dsrd.", XO
=0b10111001, Rc
=1),
616 def dsld_dsrd(fields
, XO
, Rc
):
617 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
619 # |0 |6 |11 |15 |16 |21 |23 |31 |
620 # | PO | RT | RA | RB |sm | XO |Rc |
622 (RT
, RA
, RB
, sm
) = fields
634 # decode GPR into sv extra
635 def get_extra_gpr(etype
, regmode
, field
):
636 if regmode
== 'scalar':
637 # cut into 2-bits 5-bits SS FFFFF
638 sv_extra
= field
>> 5
639 field
= field
& 0b11111
641 # cut into 5-bits 2-bits FFFFF SS
642 sv_extra
= field
& 0b11
644 return sv_extra
, field
647 # decode 3-bit CR into sv extra
648 def get_extra_cr_3bit(etype
, regmode
, field
):
649 if regmode
== 'scalar':
650 # cut into 2-bits 3-bits SS FFF
651 sv_extra
= field
>> 3
652 field
= field
& 0b111
654 # cut into 3-bits 4-bits FFF SSSS but will cut 2 zeros off later
655 sv_extra
= field
& 0b1111
657 return sv_extra
, field
661 def decode_subvl(encoding
):
662 pmap
= {'2': 0b01, '3': 0b10, '4': 0b11}
663 assert encoding
in pmap
, \
664 "encoding %s for SUBVL not recognised" % encoding
665 return pmap
[encoding
]
669 def decode_elwidth(encoding
):
670 pmap
= {'8': 0b11, '16': 0b10, '32': 0b01}
671 assert encoding
in pmap
, \
672 "encoding %s for elwidth not recognised" % encoding
673 return pmap
[encoding
]
676 # decodes predicate register encoding
677 def decode_predicate(encoding
):
688 'nl': (1, 0b001), 'ge': (1, 0b001), # same value
690 'ng': (1, 0b011), 'le': (1, 0b011), # same value
693 'so': (1, 0b110), 'un': (1, 0b110), # same value
694 'ns': (1, 0b111), 'nu': (1, 0b111), # same value
696 assert encoding
in pmap
, \
697 "encoding %s for predicate not recognised" % encoding
698 return pmap
[encoding
]
701 # decodes "Mode" in similar way to BO field (supposed to, anyway)
702 def decode_bo(encoding
):
703 pmap
= { # TODO: double-check that these are the same as Branch BO
705 'nl': 0b001, 'ge': 0b001, # same value
707 'ng': 0b011, 'le': 0b011, # same value
710 'so': 0b110, 'un': 0b110, # same value
711 'ns': 0b111, 'nu': 0b111, # same value
713 assert encoding
in pmap
, \
714 "encoding %s for BO Mode not recognised" % encoding
715 # barse-ackwards MSB0/LSB0. sigh. this would be nice to be the
716 # same as the decode_predicate() CRfield table above, but (inv,CRbit)
717 # is how it is in the spec [decode_predicate is (CRbit,inv)]
718 mapped
= pmap
[encoding
]
719 si
= SelectableInt(0, 3)
720 si
[0] = mapped
& 1 # inv
721 si
[1:3] = mapped
>> 1 # CR
725 # partial-decode fail-first mode
726 def decode_ffirst(encoding
):
727 if encoding
in ['RC1', '~RC1']:
729 return decode_bo(encoding
)
732 def decode_reg(field
, macros
=None):
735 # decode the field number. "5.v" or "3.s" or "9"
736 # and now also "*0", and "*%0". note: *NOT* to add "*%rNNN" etc.
737 # https://bugs.libre-soc.org/show_bug.cgi?id=884#c0
738 if field
.startswith(("*%", "*")):
739 if field
.startswith("*%"):
743 while field
in macros
:
744 field
= macros
[field
]
745 return int(field
), "vector" # actual register number
747 # try old convention (to be retired)
748 field
= field
.split(".")
749 regmode
= 'scalar' # default
753 elif field
[1] == 'v':
755 field
= int(field
[0]) # actual register number
756 return field
, regmode
759 def decode_imm(field
):
760 ldst_imm
= "(" in field
and field
[-1] == ')'
762 return field
[:-1].split("(")
767 def crf_extra(etype
, rname
, extra_idx
, regmode
, field
, extras
):
768 """takes a CR Field number (CR0-CR127), splits into EXTRA2/3 and v3.0
769 the scalar/vector mode (crNN.v or crNN.s) changes both the format
770 of the EXTRA2/3 encoding as well as what range of registers is possible.
771 this function can be used for both BF/BFA and BA/BB/BT by first removing
772 the bottom 2 bits of BA/BB/BT then re-instating them after encoding.
773 see https://libre-soc.org/openpower/sv/svp64/appendix/#cr_extra
776 sv_extra
, field
= get_extra_cr_3bit(etype
, regmode
, field
)
777 # now sanity-check (and shrink afterwards)
778 if etype
== 'EXTRA2':
779 # 3-bit CR Field (BF, BFA) EXTRA2 encoding
780 if regmode
== 'scalar':
781 # range is CR0-CR15 in increments of 1
782 assert (sv_extra
>> 1) == 0, \
783 "scalar CR %s cannot fit into EXTRA2 %s" % \
784 (rname
, str(extras
[extra_idx
]))
785 # all good: encode as scalar
786 sv_extra
= sv_extra
& 0b01
788 # range is CR0-CR127 in increments of 16
789 assert sv_extra
& 0b111 == 0, \
790 "vector CR %s cannot fit into EXTRA2 %s" % \
791 (rname
, str(extras
[extra_idx
]))
792 # all good: encode as vector (bit 2 set)
793 sv_extra
= 0b10 |
(sv_extra
>> 3)
795 # 3-bit CR Field (BF, BFA) EXTRA3 encoding
796 if regmode
== 'scalar':
797 # range is CR0-CR31 in increments of 1
798 assert (sv_extra
>> 2) == 0, \
799 "scalar CR %s cannot fit into EXTRA3 %s" % \
800 (rname
, str(extras
[extra_idx
]))
801 # all good: encode as scalar
802 sv_extra
= sv_extra
& 0b11
804 # range is CR0-CR127 in increments of 8
805 assert sv_extra
& 0b11 == 0, \
806 "vector CR %s cannot fit into EXTRA3 %s" % \
807 (rname
, str(extras
[extra_idx
]))
808 # all good: encode as vector (bit 3 set)
809 sv_extra
= 0b100 |
(sv_extra
>> 2)
810 return sv_extra
, field
813 def to_number(field
):
814 if field
.startswith("0x"):
816 if field
.startswith("0b"):
821 db
= Database(find_wiki_dir())
824 # decodes svp64 assembly listings and creates EXT001 svp64 prefixes
826 def __init__(self
, lst
, bigendian
=False, macros
=None):
831 self
.trans
= self
.translate(lst
)
832 self
.isa
= ISA() # reads the v3.0B pseudo-code markdown files
833 self
.svp64
= SVP64RM() # reads the svp64 Remap entries for registers
834 assert bigendian
== False, "error, bigendian not supported yet"
837 yield from self
.trans
839 def translate_one(self
, insn
, macros
=None):
842 macros
.update(self
.macros
)
845 insn_no_comments
= insn
.partition('#')[0]
846 # find first space, to get opcode
847 ls
= insn_no_comments
.split()
849 # now find opcode fields
850 fields
= ''.join(ls
[1:]).split(',')
851 mfields
= list(map(str.strip
, fields
))
852 log("opcode, fields", ls
, opcode
, mfields
)
855 for field
in mfields
:
856 fields
.append(macro_subst(macros
, field
))
857 log("opcode, fields substed", ls
, opcode
, fields
)
859 # identify if it is a special instruction
860 custom_insn_hook
= CUSTOM_INSNS
.get(opcode
)
861 if custom_insn_hook
is not None:
862 fields
= tuple(map(to_number
, fields
))
863 insn_num
= custom_insn_hook(fields
)
864 log(opcode
, bin(insn_num
))
865 yield ".long 0x%X # %s" % (insn_num
, insn
)
868 # identify if is a svp64 mnemonic
869 if not opcode
.startswith('sv.'):
870 yield insn
# unaltered
872 opcode
= opcode
[3:] # strip leading "sv"
874 # start working on decoding the svp64 op: sv.basev30Bop/vec2/mode
875 opmodes
= opcode
.split("/") # split at "/"
876 v30b_op_orig
= opmodes
.pop(0) # first is the v3.0B
877 # check instruction ends with dot
878 rc_mode
= v30b_op_orig
.endswith('.')
880 v30b_op
= v30b_op_orig
[:-1]
882 v30b_op
= v30b_op_orig
884 # look up the 32-bit op (original, with "." if it has it)
885 if v30b_op_orig
in isa
.instr
:
886 isa_instr
= isa
.instr
[v30b_op_orig
]
888 raise Exception("opcode %s of '%s' not supported" %
889 (v30b_op_orig
, insn
))
891 # look up the svp64 op, first the original (with "." if it has it)
892 if v30b_op_orig
in svp64
.instrs
:
893 rm
= svp64
.instrs
[v30b_op_orig
] # one row of the svp64 RM CSV
894 # then without the "." (if there was one)
895 elif v30b_op
in svp64
.instrs
:
896 rm
= svp64
.instrs
[v30b_op
] # one row of the svp64 RM CSV
898 raise Exception(f
"opcode {v30b_op_orig!r} of "
899 f
"{insn!r} not an svp64 instruction")
901 # get regs info e.g. "RT,RA,RB"
902 v30b_regs
= isa_instr
.regs
[0]
903 log("v3.0B op", v30b_op
, "Rc=1" if rc_mode
else '')
904 log("v3.0B regs", opcode
, v30b_regs
)
907 # right. the first thing to do is identify the ordering of
908 # the registers, by name. the EXTRA2/3 ordering is in
909 # rm['0']..rm['3'] but those fields contain the names RA, BB
910 # etc. we have to read the pseudocode to understand which
911 # reg is which in our instruction. sigh.
913 # first turn the svp64 rm into a "by name" dict, recording
914 # which position in the RM EXTRA it goes into
915 # also: record if the src or dest was a CR, for sanity-checking
916 # (elwidth overrides on CRs are banned)
917 decode
= decode_extra(rm
)
918 dest_reg_cr
, src_reg_cr
, svp64_src
, svp64_dest
= decode
920 log("EXTRA field index, src", svp64_src
)
921 log("EXTRA field index, dest", svp64_dest
)
923 # okaaay now we identify the field value (opcode N,N,N) with
924 # the pseudo-code info (opcode RT, RA, RB)
925 assert len(fields
) == len(v30b_regs
), \
926 "length of fields %s must match insn `%s` fields %s" % \
927 (str(v30b_regs
), insn
, str(fields
))
928 opregfields
= zip(fields
, v30b_regs
) # err that was easy
930 # now for each of those find its place in the EXTRA encoding
931 # note there is the possibility (for LD/ST-with-update) of
932 # RA occurring **TWICE**. to avoid it getting added to the
933 # v3.0B suffix twice, we spot it as a duplicate, here
934 extras
= OrderedDict()
935 for idx
, (field
, regname
) in enumerate(opregfields
):
936 imm
, regname
= decode_imm(regname
)
937 rtype
= get_regtype(regname
)
938 log(" idx find", rtype
, idx
, field
, regname
, imm
)
940 # probably an immediate field, append it straight
941 extras
[('imm', idx
, False)] = (idx
, field
, None, None, None)
943 extra
= svp64_src
.get(regname
, None)
944 if extra
is not None:
945 extra
= ('s', extra
, False) # not a duplicate
946 extras
[extra
] = (idx
, field
, regname
, rtype
, imm
)
947 log(" idx src", idx
, extra
, extras
[extra
])
948 dextra
= svp64_dest
.get(regname
, None)
949 log("regname in", regname
, dextra
)
950 if dextra
is not None:
951 is_a_duplicate
= extra
is not None # duplicate spotted
952 dextra
= ('d', dextra
, is_a_duplicate
)
953 extras
[dextra
] = (idx
, field
, regname
, rtype
, imm
)
954 log(" idx dst", idx
, extra
, extras
[dextra
])
956 # great! got the extra fields in their associated positions:
957 # also we know the register type. now to create the EXTRA encodings
958 etype
= rm
['Etype'] # Extra type: EXTRA3/EXTRA2
959 ptype
= rm
['Ptype'] # Predication type: Twin / Single
962 for extra_idx
, (idx
, field
, rname
, rtype
, iname
) in extras
.items():
963 # is it a field we don't alter/examine? if so just put it
966 v30b_newfields
.append(field
)
969 # identify if this is a ld/st immediate(reg) thing
970 ldst_imm
= "(" in field
and field
[-1] == ')'
972 immed
, field
= field
[:-1].split("(")
974 field
, regmode
= decode_reg(field
, macros
=macros
)
975 log(" ", extra_idx
, rname
, rtype
,
976 regmode
, iname
, field
, end
=" ")
978 # see Mode field https://libre-soc.org/openpower/sv/svp64/
979 # XXX TODO: the following is a bit of a laborious repeated
980 # mess, which could (and should) easily be parameterised.
981 # XXX also TODO: the LD/ST modes which are different
982 # https://libre-soc.org/openpower/sv/ldst/
984 # rright. SVP64 register numbering is from 0 to 127
985 # for GPRs, FPRs *and* CR Fields, where for v3.0 the GPRs and RPFs
986 # are 0-31 and CR Fields are only 0-7. the SVP64 RM "Extra"
987 # area is used to extend the numbering from the 32-bit
988 # instruction, and also to record whether the register
989 # is scalar or vector. on a per-operand basis. this
990 # results in a slightly finnicky encoding: here we go...
992 # encode SV-GPR and SV-FPR field into extra, v3.0field
993 if rtype
in ['GPR', 'FPR']:
994 sv_extra
, field
= get_extra_gpr(etype
, regmode
, field
)
995 # now sanity-check. EXTRA3 is ok, EXTRA2 has limits
996 # (and shrink to a single bit if ok)
997 if etype
== 'EXTRA2':
998 if regmode
== 'scalar':
999 # range is r0-r63 in increments of 1
1000 assert (sv_extra
>> 1) == 0, \
1001 "scalar GPR %s cannot fit into EXTRA2 %s" % \
1002 (rname
, str(extras
[extra_idx
]))
1003 # all good: encode as scalar
1004 sv_extra
= sv_extra
& 0b01
1006 # range is r0-r127 in increments of 2 (r0 r2 ... r126)
1007 assert sv_extra
& 0b01 == 0, \
1008 "%s: vector field %s cannot fit " \
1009 "into EXTRA2 %s" % \
1010 (insn
, rname
, str(extras
[extra_idx
]))
1011 # all good: encode as vector (bit 2 set)
1012 sv_extra
= 0b10 |
(sv_extra
>> 1)
1013 elif regmode
== 'vector':
1014 # EXTRA3 vector bit needs marking
1017 # encode SV-CR 3-bit field into extra, v3.0field.
1018 # 3-bit is for things like BF and BFA
1019 elif rtype
== 'CR_3bit':
1020 sv_extra
, field
= crf_extra(etype
, rname
, extra_idx
,
1021 regmode
, field
, extras
)
1023 # encode SV-CR 5-bit field into extra, v3.0field
1024 # 5-bit is for things like BA BB BC BT etc.
1025 # *sigh* this is the same as 3-bit except the 2 LSBs of the
1026 # 5-bit field are passed through unaltered.
1027 elif rtype
== 'CR_5bit':
1028 cr_subfield
= field
& 0b11 # record bottom 2 bits for later
1029 field
= field
>> 2 # strip bottom 2 bits
1030 # use the exact same 3-bit function for the top 3 bits
1031 sv_extra
, field
= crf_extra(etype
, rname
, extra_idx
,
1032 regmode
, field
, extras
)
1033 # reconstruct the actual 5-bit CR field (preserving the
1034 # bottom 2 bits, unaltered)
1035 field
= (field
<< 2) | cr_subfield
1038 raise Exception("no type match: %s" % rtype
)
1040 # capture the extra field info
1041 log("=>", "%5s" % bin(sv_extra
), field
)
1042 extras
[extra_idx
] = sv_extra
1044 # append altered field value to v3.0b, differs for LDST
1045 # note that duplicates are skipped e.g. EXTRA2 contains
1046 # *BOTH* s:RA *AND* d:RA which happens on LD/ST-with-update
1047 srcdest
, idx
, duplicate
= extra_idx
1048 if duplicate
: # skip adding to v3.0b fields, already added
1051 v30b_newfields
.append(("%s(%s)" % (immed
, str(field
))))
1053 v30b_newfields
.append(str(field
))
1055 log("new v3.0B fields", v30b_op
, v30b_newfields
)
1056 log("extras", extras
)
1058 # rright. now we have all the info. start creating SVP64 instruction.
1059 svp64_insn
= SVP64Instruction
.pair(prefix
=0, suffix
=0)
1060 svp64_prefix
= svp64_insn
.prefix
1061 svp64_rm
= svp64_insn
.prefix
.rm
1063 # begin with EXTRA fields
1064 for idx
, sv_extra
in extras
.items():
1070 srcdest
, idx
, duplicate
= idx
1071 if etype
== 'EXTRA2':
1072 svp64_rm
.extra2
[idx
] = sv_extra
1074 svp64_rm
.extra3
[idx
] = sv_extra
1076 # identify if the op is a LD/ST.
1077 # see https://libre-soc.org/openpower/sv/ldst/
1078 is_ldst
= rm
['mode'] in [ 'LDST_IDX', 'LDST_IMM']
1079 is_ldst_idx
= rm
['mode'] == 'LDST_IDX'
1080 is_ld
= v30b_op
.startswith("l") and is_ldst
1081 is_st
= v30b_op
.startswith("s") and is_ldst
1083 # branch-conditional detection
1084 is_bc
= rm
['mode'] == 'BRANCH'
1088 pmask
= 0 # bits 1-3
1089 destwid
= 0 # bits 4-5
1090 srcwid
= 0 # bits 6-7
1091 subvl
= 0 # bits 8-9
1092 smask
= 0 # bits 16-18 but only for twin-predication
1093 mode
= 0 # bits 19-23
1095 mask_m_specified
= False
1105 reverse_gear
= False
1106 mapreduce_crm
= False
1116 # ok let's start identifying opcode augmentation fields
1117 for encmode
in opmodes
:
1118 # predicate mask (src and dest)
1119 if encmode
.startswith("m="):
1121 pmmode
, pmask
= decode_predicate(encmode
[2:])
1122 smmode
, smask
= pmmode
, pmask
1124 mask_m_specified
= True
1125 # predicate mask (dest)
1126 elif encmode
.startswith("dm="):
1128 pmmode
, pmask
= decode_predicate(encmode
[3:])
1131 # predicate mask (src, twin-pred)
1132 elif encmode
.startswith("sm="):
1134 smmode
, smask
= decode_predicate(encmode
[3:])
1138 elif encmode
.startswith("vec"):
1139 subvl
= decode_subvl(encmode
[3:])
1140 # elwidth (both src and dest, like mask)
1141 elif encmode
.startswith("w="):
1142 destwid
= decode_elwidth(encmode
[2:])
1143 srcwid
= decode_elwidth(encmode
[2:])
1145 elif encmode
.startswith("dw="):
1146 destwid
= decode_elwidth(encmode
[3:])
1148 elif encmode
.startswith("sw="):
1149 srcwid
= decode_elwidth(encmode
[3:])
1150 # element-strided LD/ST
1151 elif encmode
== 'els':
1153 # in indexed mode, set sv_mode=0b01
1157 elif encmode
== 'sats':
1158 assert sv_mode
is None
1161 elif encmode
== 'satu':
1162 assert sv_mode
is None
1166 elif encmode
== 'zz': # TODO, a lot more checking on legality
1167 dst_zero
= 1 # NOT on cr_ops, that's RM[6]
1169 elif encmode
== 'sz':
1171 elif encmode
== 'dz':
1174 elif encmode
.startswith("ff="):
1175 assert sv_mode
is None
1177 failfirst
= decode_ffirst(encmode
[3:])
1178 assert sea
is False, "cannot use failfirst with signed-address"
1179 # predicate-result, interestingly same as fail-first
1180 elif encmode
.startswith("pr="):
1181 assert sv_mode
is None
1183 predresult
= decode_ffirst(encmode
[3:])
1184 # map-reduce mode, reverse-gear
1185 elif encmode
== 'mrr':
1186 assert sv_mode
is None
1191 elif encmode
== 'mr':
1192 assert sv_mode
is None
1195 elif encmode
== 'crm': # CR on map-reduce
1196 assert sv_mode
is None
1198 mapreduce_crm
= True
1199 elif encmode
== 'vli':
1200 assert sv_mode
== 0b01 # only allow ff mode
1202 elif encmode
== 'sea':
1205 assert failfirst
is False, "cannot use ffirst+signed-address"
1207 if encmode
== 'all':
1208 svp64_rm
.branch
.ALL
= 1
1209 elif encmode
== 'snz':
1210 svp64_rm
.branch
.sz
= 1
1211 svp64_rm
.branch
.SNZ
= 1
1212 elif encmode
== 'sl':
1213 svp64_rm
.branch
.SL
= 1
1214 elif encmode
== 'slu':
1215 svp64_rm
.branch
.SLu
= 1
1216 elif encmode
== 'lru':
1217 svp64_rm
.branch
.LRu
= 1
1218 elif encmode
== 'vs':
1219 svp64_rm
.branch
.VLS
= 1
1220 elif encmode
== 'vsi':
1221 svp64_rm
.branch
.VLS
= 1
1222 svp64_rm
.branch
.vls
.VLi
= 1
1223 elif encmode
== 'vsb':
1224 svp64_rm
.branch
.VLS
= 1
1225 svp64_rm
.branch
.vls
.VSb
= 1
1226 elif encmode
== 'vsbi':
1227 svp64_rm
.branch
.VLS
= 1
1228 svp64_rm
.branch
.vls
.VSb
= 1
1229 svp64_rm
.branch
.vls
.VLi
= 1
1230 elif encmode
== 'ctr':
1231 svp64_rm
.branch
.CTR
= 1
1232 elif encmode
== 'cti':
1233 svp64_rm
.branch
.CTR
= 1
1234 svp64_rm
.branch
.ctr
.CTi
= 1
1236 raise AssertionError("unknown encmode %s" % encmode
)
1238 raise AssertionError("unknown encmode %s" % encmode
)
1240 # sanity check if dz/zz used in branch-mode
1241 if is_bc
and dst_zero
:
1242 raise AssertionError("dz/zz not supported in branch, use 'sz'")
1244 # check sea *after* all qualifiers are evaluated
1246 assert sv_mode
in (None, 0b00, 0b01)
1249 # since m=xx takes precedence (overrides) sm=xx and dm=xx,
1250 # treat them as mutually exclusive
1251 if mask_m_specified
:
1252 assert not has_smask
,\
1253 "cannot have both source-mask and predicate mask"
1254 assert not has_pmask
,\
1255 "cannot have both dest-mask and predicate mask"
1256 # since the default is INT predication (ALWAYS), if you
1257 # specify one CR mask, you must specify both, to avoid
1258 # mixing INT and CR reg types
1259 if has_pmask
and pmmode
== 1:
1261 "need explicit source-mask in CR twin predication"
1262 if has_smask
and smmode
== 1:
1264 "need explicit dest-mask in CR twin predication"
1265 # sanity-check that 2Pred mask is same mode
1266 if has_pmask
and has_smask
:
1267 assert smmode
== pmmode
, \
1268 "predicate masks %s and %s must be same reg type" % \
1271 # sanity-check that twin-predication mask only specified in 2P mode
1273 assert not has_smask
, \
1274 "source-mask can only be specified on Twin-predicate ops"
1275 assert not has_pmask
, \
1276 "dest-mask can only be specified on Twin-predicate ops"
1278 # construct the mode field, doing sanity-checking along the way
1280 assert has_smask
or mask_m_specified
, \
1281 "src zeroing requires a source predicate"
1283 assert has_pmask
or mask_m_specified
, \
1284 "dest zeroing requires a dest predicate"
1286 # okaaay, so there are 4 different modes, here, which will be
1287 # partly-merged-in: is_ldst is merged in with "normal", but
1288 # is_bc is so different it's done separately. likewise is_cr
1289 # (when it is done). here are the maps:
1291 # for "normal" arithmetic: https://libre-soc.org/openpower/sv/normal/
1293 | 0-1 | 2 | 3 4 | description |
1294 | --- | --- |---------|-------------------------- |
1295 | 00 | 0 | dz sz | simple mode |
1296 | 00 | 1 | 0 RG | scalar reduce mode (mapreduce) |
1297 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
1298 | 01 | inv | VLi RC1 | Rc=0: ffirst z/nonz |
1299 | 10 | N | dz sz | sat mode: N=0/1 u/s |
1300 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
1301 | 11 | inv | zz RC1 | Rc=0: pred-result z/nonz |
1304 # https://libre-soc.org/openpower/sv/ldst/
1305 # for LD/ST-immediate:
1307 | 0-1 | 2 | 3 4 | description |
1308 | --- | --- |---------|--------------------------- |
1309 | 00 | 0 | zz els | normal mode |
1310 | 00 | 1 | / / | reserved |
1311 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
1312 | 01 | inv | els RC1 | Rc=0: ffirst z/nonz |
1313 | 10 | N | zz els | sat mode: N=0/1 u/s |
1314 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
1315 | 11 | inv | els RC1 | Rc=0: pred-result z/nonz |
1318 # for LD/ST-indexed (RA+RB):
1320 | 0-1 | 2 | 3 4 | description |
1321 | --- | --- |---------|----------------------------- |
1322 | 00 | SEA | dz sz | normal mode |
1323 | 01 | SEA | dz sz | strided (scalar only source) |
1324 | 10 | N | dz sz | sat mode: N=0/1 u/s |
1325 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
1326 | 11 | inv | dz RC1 | Rc=0: pred-result z/nonz |
1329 # and leaving out branches and cr_ops for now because they're
1331 """ TODO branches and cr_ops
1335 sv_mode
= int(svp64_rm
.mode
[0, 1])
1337 svp64_rm
.branch
.sz
= 1
1340 ######################################
1341 # "element-strided" mode, ldst_idx
1342 if sv_mode
== 0b01 and is_ldst_idx
:
1343 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
1344 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1345 mode |
= sea
<< SVP64MODE
.SEA
# el-strided
1347 ######################################
1349 elif sv_mode
is None:
1350 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
1351 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1353 # TODO: for now, LD/ST-indexed is ignored.
1354 mode |
= ldst_elstride
<< SVP64MODE
.ELS_NORMAL
# el-strided
1356 # TODO, reduce and subvector mode
1357 # 00 1 dz CRM reduce mode (mapreduce), SUBVL=1
1358 # 00 1 SVM CRM subvector reduce mode, SUBVL>1
1362 ######################################
1364 elif sv_mode
== 0b00:
1365 mode |
= (0b1 << SVP64MODE
.REDUCE
) # sets mapreduce
1366 assert dst_zero
== 0, "dest-zero not allowed in mapreduce mode"
1368 mode |
= (0b1 << SVP64MODE
.RG
) # sets Reverse-gear mode
1370 mode |
= (0b1 << SVP64MODE
.CRM
) # sets CRM mode
1371 assert rc_mode
, "CRM only allowed when Rc=1"
1372 # bit of weird encoding to jam zero-pred or SVM mode in.
1373 # SVM mode can be enabled only when SUBVL=2/3/4 (vec2/3/4)
1375 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1377 ######################################
1379 elif sv_mode
== 0b01:
1380 assert src_zero
== 0, "dest-zero not allowed in failfirst mode"
1381 if failfirst
== 'RC1':
1382 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
1383 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1384 assert rc_mode
== False, "ffirst RC1 only ok when Rc=0"
1385 elif failfirst
== '~RC1':
1386 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
1387 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1388 mode |
= (0b1 << SVP64MODE
.INV
) # ... with inversion
1389 assert rc_mode
== False, "ffirst RC1 only ok when Rc=0"
1391 assert dst_zero
== 0, "dst-zero not allowed in ffirst BO"
1392 assert rc_mode
, "ffirst BO only possible when Rc=1"
1393 mode |
= (failfirst
<< SVP64MODE
.BO_LSB
) # set BO
1395 ######################################
1396 # "saturation" modes
1397 elif sv_mode
== 0b10:
1398 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
1399 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1400 mode |
= (saturation
<< SVP64MODE
.N
) # signed/us saturation
1402 ######################################
1403 # "predicate-result" modes. err... code-duplication from ffirst
1404 elif sv_mode
== 0b11:
1405 assert src_zero
== 0, "dest-zero not allowed in predresult mode"
1406 if predresult
== 'RC1':
1407 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
1408 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1409 assert rc_mode
== False, "pr-mode RC1 only ok when Rc=0"
1410 elif predresult
== '~RC1':
1411 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
1412 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1413 mode |
= (0b1 << SVP64MODE
.INV
) # ... with inversion
1414 assert rc_mode
== False, "pr-mode RC1 only ok when Rc=0"
1416 assert dst_zero
== 0, "dst-zero not allowed in pr-mode BO"
1417 assert rc_mode
, "pr-mode BO only possible when Rc=1"
1418 mode |
= (predresult
<< SVP64MODE
.BO_LSB
) # set BO
1420 # whewww.... modes all done :)
1421 # now put into svp64_rm, but respect MSB0 order
1423 mode |
= (0b1<<SVP64MODE
.MOD2_LSB
)
1425 mode |
= (0b1<<SVP64MODE
.MOD2_MSB
)
1428 mode |
= (0b1 << SVP64MODE
.SEA
)
1431 svp64_rm
.mode
= mode
# mode: bits 19-23
1433 svp64_rm
.normal
.ffrc0
.VLi
= 1
1435 # put in predicate masks into svp64_rm
1437 svp64_rm
.smask
= smask
# source pred: bits 16-18
1439 # put in elwidths unless bc
1440 svp64_rm
.ewsrc
= srcwid
# srcwid: bits 6-7
1441 svp64_rm
.elwidth
= destwid
# destwid: bits 4-5
1443 svp64_rm
.mmode
= mmode
# mask mode: bit 0
1444 svp64_rm
.mask
= pmask
# 1-pred: bits 1-3
1445 svp64_rm
.subvl
= subvl
# and subvl: bits 8-9
1447 # nice debug printout. (and now for something completely different)
1448 # https://youtu.be/u0WOIwlXE9g?t=146
1449 svp64_rm_value
= int(svp64_rm
)
1450 log("svp64_rm", hex(svp64_rm_value
), bin(svp64_rm_value
))
1451 log(" mmode 0 :", bin(mmode
))
1452 log(" pmask 1-3 :", bin(pmask
))
1453 log(" dstwid 4-5 :", bin(destwid
))
1454 log(" srcwid 6-7 :", bin(srcwid
))
1455 log(" subvl 8-9 :", bin(subvl
))
1456 log(" mode 19-23:", bin(svp64_rm
.mode
))
1457 offs
= 2 if etype
== 'EXTRA2' else 3 # 2 or 3 bits
1458 for idx
, sv_extra
in extras
.items():
1463 srcdest
, idx
, duplicate
= idx
1464 start
= (10+idx
*offs
)
1465 end
= start
+ offs
-1
1466 log(" extra%d %2d-%2d:" % (idx
, start
, end
),
1469 log(" smask 16-17:", bin(smask
))
1472 # update prefix PO and ID (aka PID)
1473 svp64_prefix
.po
= 0x1
1474 svp64_prefix
.id = 0b11
1476 # fiinally yield the svp64 prefix and the thingy. v3.0b opcode
1477 rc
= '.' if rc_mode
else ''
1478 yield ".long 0x%08x" % int(svp64_prefix
)
1479 log(v30b_op
, v30b_newfields
)
1481 v30b_op_rc
= v30b_op
1482 if not v30b_op
.endswith('.'):
1486 # FIXME(lkcl): should sv.svstep be like svstep?
1487 if v30b_op_rc
in ("svstep", "svstep."):
1488 # compensate for `SVi -= 1` in svstep()
1489 v30b_newfields
[1] = str(int(v30b_newfields
[1]) + 1)
1491 custom_insn_hook
= CUSTOM_INSNS
.get(v30b_op_rc
)
1492 if custom_insn_hook
is not None:
1493 fields
= tuple(map(to_number
, v30b_newfields
))
1494 insn_num
= custom_insn_hook(fields
)
1495 log(opcode
, bin(insn_num
))
1496 yield ".long 0x%X # %s" % (insn_num
, insn
)
1498 # argh, sv.fmadds etc. need to be done manually
1499 elif v30b_op
== 'ffmadds':
1500 opcode
= 59 << (32-6) # bits 0..6 (MSB0)
1501 opcode |
= int(v30b_newfields
[0]) << (32-11) # FRT
1502 opcode |
= int(v30b_newfields
[1]) << (32-16) # FRA
1503 opcode |
= int(v30b_newfields
[2]) << (32-21) # FRB
1504 opcode |
= int(v30b_newfields
[3]) << (32-26) # FRC
1505 opcode |
= 0b00101 << (32-31) # bits 26-30
1507 opcode |
= 1 # Rc, bit 31.
1508 yield ".long 0x%x" % opcode
1509 # argh, sv.fdmadds need to be done manually
1510 elif v30b_op
== 'fdmadds':
1511 opcode
= 59 << (32-6) # bits 0..6 (MSB0)
1512 opcode |
= int(v30b_newfields
[0]) << (32-11) # FRT
1513 opcode |
= int(v30b_newfields
[1]) << (32-16) # FRA
1514 opcode |
= int(v30b_newfields
[2]) << (32-21) # FRB
1515 opcode |
= int(v30b_newfields
[3]) << (32-26) # FRC
1516 opcode |
= 0b11011 << (32-31) # bits 26-30
1518 opcode |
= 1 # Rc, bit 31.
1519 yield ".long 0x%x" % opcode
1520 # argh, sv.ffadds etc. need to be done manually
1521 elif v30b_op
== 'ffadds':
1522 opcode
= 59 << (32-6) # bits 0..6 (MSB0)
1523 opcode |
= int(v30b_newfields
[0]) << (32-11) # FRT
1524 opcode |
= int(v30b_newfields
[1]) << (32-16) # FRA
1525 opcode |
= int(v30b_newfields
[2]) << (32-21) # FRB
1526 opcode |
= 0b1111100000 << (32-31) # bits 21-30
1528 opcode |
= 1 # Rc, bit 31.
1529 yield ".long 0x%x" % opcode
1531 if not v30b_op
.endswith('.'):
1533 yield "%s %s" % (v30b_op
, ", ".join(v30b_newfields
))
1534 for (name
, span
) in svp64_insn
.traverse("SVP64"):
1535 value
= svp64_insn
.storage
[span
]
1536 log(name
, f
"{value.value:0{value.bits}b}", span
)
1537 log("new v3.0B fields", v30b_op
, v30b_newfields
)
1539 def translate(self
, lst
):
1541 yield from self
.translate_one(insn
)
1544 def macro_subst(macros
, txt
):
1546 log("subst", txt
, macros
)
1549 for macro
, value
in macros
.items():
1552 replaced
= txt
.replace(macro
, value
)
1553 log("macro", txt
, "replaced", replaced
, macro
, value
)
1556 toreplace
= '%s.s' % macro
1557 if toreplace
== txt
:
1559 replaced
= txt
.replace(toreplace
, "%s.s" % value
)
1560 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1563 toreplace
= '%s.v' % macro
1564 if toreplace
== txt
:
1566 replaced
= txt
.replace(toreplace
, "%s.v" % value
)
1567 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1570 toreplace
= '*%s' % macro
1571 if toreplace
in txt
:
1573 replaced
= txt
.replace(toreplace
, '*%s' % value
)
1574 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1577 toreplace
= '(%s)' % macro
1578 if toreplace
in txt
:
1580 replaced
= txt
.replace(toreplace
, '(%s)' % value
)
1581 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1584 log(" processed", txt
)
1592 if not line
[0].isspace():
1600 # get an input file and an output file
1604 outfile
= sys
.stdout
1605 # read the whole lot in advance in case of in-place
1606 lines
= list(infile
.readlines())
1607 elif len(args
) != 2:
1608 print("pysvp64asm [infile | -] [outfile | -]", file=sys
.stderr
)
1614 infile
= open(args
[0], "r")
1615 # read the whole lot in advance in case of in-place overwrite
1616 lines
= list(infile
.readlines())
1619 outfile
= sys
.stdout
1621 outfile
= open(args
[1], "w")
1623 # read the line, look for custom insn, process it
1624 macros
= {} # macros which start ".set"
1627 op
= line
.split("#")[0].strip()
1629 if op
.startswith(".set"):
1630 macro
= op
[4:].split(",")
1631 (macro
, value
) = map(str.strip
, macro
)
1632 macros
[macro
] = value
1633 if not op
.startswith('sv.') and not op
.startswith(tuple(CUSTOM_INSNS
)):
1637 (ws
, line
) = get_ws(line
)
1638 lst
= isa
.translate_one(op
, macros
)
1639 lst
= '; '.join(lst
)
1640 outfile
.write("%s%s # %s\n" % (ws
, lst
, op
))
1643 if __name__
== '__main__':
1644 lst
= ['slw 3, 1, 4',
1647 'sv.cmpi 5, 1, 3, 2',
1649 'sv.isel 64.v, 3, 2, 65.v',
1650 'sv.setb/dm=r3/sm=1<<r3 5, 31',
1651 'sv.setb/m=r3 5, 31',
1652 'sv.setb/vec2 5, 31',
1653 'sv.setb/sw=8/ew=16 5, 31',
1654 'sv.extsw./ff=eq 5, 31',
1655 'sv.extsw./satu/sz/dz/sm=r3/dm=r3 5, 31',
1656 'sv.extsw./pr=eq 5.v, 31',
1657 'sv.add. 5.v, 2.v, 1.v',
1658 'sv.add./m=r3 5.v, 2.v, 1.v',
1661 'sv.stw 5.v, 4(1.v)',
1662 'sv.ld 5.v, 4(1.v)',
1663 'setvl. 2, 3, 4, 0, 1, 1',
1664 'sv.setvl. 2, 3, 4, 0, 1, 1',
1667 "sv.stfsu 0.v, 16(4.v)",
1670 "sv.stfsu/els 0.v, 16(4)",
1673 'sv.add./mr 5.v, 2.v, 1.v',
1675 macros
= {'win2': '50', 'win': '60'}
1677 'sv.addi win2.v, win.v, -1',
1678 'sv.add./mrr 5.v, 2.v, 1.v',
1679 #'sv.lhzsh 5.v, 11(9.v), 15',
1680 #'sv.lwzsh 5.v, 11(9.v), 15',
1681 'sv.ffmadds 6.v, 2.v, 4.v, 6.v',
1684 #'sv.fmadds 0.v, 8.v, 16.v, 4.v',
1685 #'sv.ffadds 0.v, 8.v, 4.v',
1686 'svremap 11, 0, 1, 2, 3, 2, 1',
1687 'svshape 8, 1, 1, 1, 0',
1688 'svshape 8, 1, 1, 1, 1',
1691 #'sv.lfssh 4.v, 11(8.v), 15',
1692 #'sv.lwzsh 4.v, 11(8.v), 15',
1693 #'sv.svstep. 2.v, 4, 0',
1694 #'sv.fcfids. 48.v, 64.v',
1695 'sv.fcoss. 80.v, 0.v',
1696 'sv.fcoss. 20.v, 0.v',
1699 'sv.bc/all 3,12,192',
1700 'sv.bclr/vsbi 3,81.v,192',
1701 'sv.ld 5.v, 4(1.v)',
1702 'sv.svstep. 2.v, 4, 0',
1713 'svindex 0,0,1,0,0,0,0',
1716 'sv.svstep./m=r3 2.v, 4, 0',
1717 'ternlogi 0,0,0,0x5',
1729 'sv.andi. *80, *80, 1',
1730 'sv.ffmadds. 6.v, 2.v, 4.v, 6.v', # incorrectly inserted 32-bit op
1731 'sv.ffmadds 6.v, 2.v, 4.v, 6.v', # correctly converted to .long
1732 'svshape2 8, 1, 31, 7, 1, 1',
1733 'sv.ld 5.v, 4(1.v)',
1734 'sv.stw 5.v, 4(1.v)',
1735 'sv.bc/all 3,12,192',
1738 isa
= SVP64Asm(lst
, macros
=macros
)
1739 log("list:\n", "\n\t".join(list(isa
)))
1740 # running svp64.py is designed to test hard-coded lists
1741 # (above) - which strictly speaking should all be unit tests.
1742 # if you need to actually do assembler translation at the
1743 # commandline use "pysvp64asm" - see setup.py
1744 # XXX NO. asm_process()