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("maddedu", XO
=50),
569 _insn("maddedus", XO
=57),
570 _insn("divmod2du", XO
=58),
571 _insn("pcdec.", XO
=56),
573 def va_form(fields
, XO
):
574 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
576 # |0 |6 |11 |16 |21 |26 |
577 # | PO | RT | RA | RB | RC | XO |
579 (RT
, RA
, RB
, RC
) = fields
591 _insn("shadd", PO
=22, XO
=0b01101110, Rc
=0),
592 _insn("shadd.", PO
=22, XO
=0b01101110, Rc
=1),
593 _insn("shadduw", PO
=22, XO
=0b11101110, Rc
=0),
594 _insn("shadduw.", PO
=22, XO
=0b11101110, Rc
=1),
596 def Z23(fields
, PO
, XO
, Rc
):
597 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
599 # |0 |6 |11 |15 |16 |21 |23 |31 |
600 # | PO | RT | RA | RB |sm | XO |Rc |
601 (RT
, RA
, RB
, sm
) = fields
614 _insn("dsld", XO
=26, Rc
=0), # minor_4=52 (26<<1 | Rc=0)
615 _insn("dsld.", XO
=26, Rc
=1), # minor_4=53 (26<<1 | Rc=1)
616 _insn("dsrd", XO
=27, Rc
=0), # minor_4=54 (27<<1 | Rc=0)
617 _insn("dsrd.", XO
=27, Rc
=1), # minor_4=55 (27<<1 | Rc=1)
619 def dsld_dsrd(fields
, XO
, Rc
):
620 # XXX WARNING THESE ARE NOT APPROVED BY OPF ISA WG
622 # |0 |6 |11 |16 |21 |26 |31|
623 # | PO | RT | RA | RB | RC | XO |Rc|
625 (RT
, RA
, RB
, RC
) = fields
637 # decode GPR into sv extra
638 def get_extra_gpr(etype
, regmode
, field
):
639 if regmode
== 'scalar':
640 # cut into 2-bits 5-bits SS FFFFF
641 sv_extra
= field
>> 5
642 field
= field
& 0b11111
644 # cut into 5-bits 2-bits FFFFF SS
645 sv_extra
= field
& 0b11
647 return sv_extra
, field
650 # decode 3-bit CR into sv extra
651 def get_extra_cr_3bit(etype
, regmode
, field
):
652 if regmode
== 'scalar':
653 # cut into 2-bits 3-bits SS FFF
654 sv_extra
= field
>> 3
655 field
= field
& 0b111
657 # cut into 3-bits 4-bits FFF SSSS but will cut 2 zeros off later
658 sv_extra
= field
& 0b1111
660 return sv_extra
, field
664 def decode_subvl(encoding
):
665 pmap
= {'2': 0b01, '3': 0b10, '4': 0b11}
666 assert encoding
in pmap
, \
667 "encoding %s for SUBVL not recognised" % encoding
668 return pmap
[encoding
]
672 def decode_elwidth(encoding
):
673 pmap
= {'8': 0b11, '16': 0b10, '32': 0b01}
674 assert encoding
in pmap
, \
675 "encoding %s for elwidth not recognised" % encoding
676 return pmap
[encoding
]
679 # decodes predicate register encoding
680 def decode_predicate(encoding
):
691 'nl': (1, 0b001), 'ge': (1, 0b001), # same value
693 'ng': (1, 0b011), 'le': (1, 0b011), # same value
696 'so': (1, 0b110), 'un': (1, 0b110), # same value
697 'ns': (1, 0b111), 'nu': (1, 0b111), # same value
699 assert encoding
in pmap
, \
700 "encoding %s for predicate not recognised" % encoding
701 return pmap
[encoding
]
704 # decodes "Mode" in similar way to BO field (supposed to, anyway)
705 def decode_bo(encoding
):
706 pmap
= { # TODO: double-check that these are the same as Branch BO
708 'nl': 0b001, 'ge': 0b001, # same value
710 'ng': 0b011, 'le': 0b011, # same value
713 'so': 0b110, 'un': 0b110, # same value
714 'ns': 0b111, 'nu': 0b111, # same value
716 assert encoding
in pmap
, \
717 "encoding %s for BO Mode not recognised" % encoding
718 # barse-ackwards MSB0/LSB0. sigh. this would be nice to be the
719 # same as the decode_predicate() CRfield table above, but (inv,CRbit)
720 # is how it is in the spec [decode_predicate is (CRbit,inv)]
721 mapped
= pmap
[encoding
]
722 si
= SelectableInt(0, 3)
723 si
[0] = mapped
& 1 # inv
724 si
[1:3] = mapped
>> 1 # CR
728 # partial-decode fail-first mode
729 def decode_ffirst(encoding
):
730 if encoding
in ['RC1', '~RC1']:
732 return decode_bo(encoding
)
735 def decode_reg(field
, macros
=None):
738 # decode the field number. "5.v" or "3.s" or "9"
739 # and now also "*0", and "*%0". note: *NOT* to add "*%rNNN" etc.
740 # https://bugs.libre-soc.org/show_bug.cgi?id=884#c0
741 if field
.startswith(("*%", "*")):
742 if field
.startswith("*%"):
746 while field
in macros
:
747 field
= macros
[field
]
748 return int(field
), "vector" # actual register number
750 # try old convention (to be retired)
751 field
= field
.split(".")
752 regmode
= 'scalar' # default
756 elif field
[1] == 'v':
758 field
= int(field
[0]) # actual register number
759 return field
, regmode
762 def decode_imm(field
):
763 ldst_imm
= "(" in field
and field
[-1] == ')'
765 return field
[:-1].split("(")
770 def crf_extra(etype
, rname
, extra_idx
, regmode
, field
, extras
):
771 """takes a CR Field number (CR0-CR127), splits into EXTRA2/3 and v3.0
772 the scalar/vector mode (crNN.v or crNN.s) changes both the format
773 of the EXTRA2/3 encoding as well as what range of registers is possible.
774 this function can be used for both BF/BFA and BA/BB/BT by first removing
775 the bottom 2 bits of BA/BB/BT then re-instating them after encoding.
776 see https://libre-soc.org/openpower/sv/svp64/appendix/#cr_extra
779 sv_extra
, field
= get_extra_cr_3bit(etype
, regmode
, field
)
780 # now sanity-check (and shrink afterwards)
781 if etype
== 'EXTRA2':
782 # 3-bit CR Field (BF, BFA) EXTRA2 encoding
783 if regmode
== 'scalar':
784 # range is CR0-CR15 in increments of 1
785 assert (sv_extra
>> 1) == 0, \
786 "scalar CR %s cannot fit into EXTRA2 %s" % \
787 (rname
, str(extras
[extra_idx
]))
788 # all good: encode as scalar
789 sv_extra
= sv_extra
& 0b01
791 # range is CR0-CR127 in increments of 16
792 assert sv_extra
& 0b111 == 0, \
793 "vector CR %s cannot fit into EXTRA2 %s" % \
794 (rname
, str(extras
[extra_idx
]))
795 # all good: encode as vector (bit 2 set)
796 sv_extra
= 0b10 |
(sv_extra
>> 3)
798 # 3-bit CR Field (BF, BFA) EXTRA3 encoding
799 if regmode
== 'scalar':
800 # range is CR0-CR31 in increments of 1
801 assert (sv_extra
>> 2) == 0, \
802 "scalar CR %s cannot fit into EXTRA3 %s" % \
803 (rname
, str(extras
[extra_idx
]))
804 # all good: encode as scalar
805 sv_extra
= sv_extra
& 0b11
807 # range is CR0-CR127 in increments of 8
808 assert sv_extra
& 0b11 == 0, \
809 "vector CR %s cannot fit into EXTRA3 %s" % \
810 (rname
, str(extras
[extra_idx
]))
811 # all good: encode as vector (bit 3 set)
812 sv_extra
= 0b100 |
(sv_extra
>> 2)
813 return sv_extra
, field
816 def to_number(field
):
817 if field
.startswith("0x"):
819 if field
.startswith("0b"):
824 db
= Database(find_wiki_dir())
827 # decodes svp64 assembly listings and creates EXT001 svp64 prefixes
829 def __init__(self
, lst
, bigendian
=False, macros
=None):
834 self
.trans
= self
.translate(lst
)
835 self
.isa
= ISA() # reads the v3.0B pseudo-code markdown files
836 self
.svp64
= SVP64RM() # reads the svp64 Remap entries for registers
837 assert bigendian
== False, "error, bigendian not supported yet"
840 yield from self
.trans
842 def translate_one(self
, insn
, macros
=None):
845 macros
.update(self
.macros
)
848 insn_no_comments
= insn
.partition('#')[0]
849 # find first space, to get opcode
850 ls
= insn_no_comments
.split()
852 # now find opcode fields
853 fields
= ''.join(ls
[1:]).split(',')
854 mfields
= list(map(str.strip
, fields
))
855 log("opcode, fields", ls
, opcode
, mfields
)
858 for field
in mfields
:
859 fields
.append(macro_subst(macros
, field
))
860 log("opcode, fields substed", ls
, opcode
, fields
)
862 # identify if it is a special instruction
863 custom_insn_hook
= CUSTOM_INSNS
.get(opcode
)
864 if custom_insn_hook
is not None:
865 fields
= tuple(map(to_number
, fields
))
866 insn_num
= custom_insn_hook(fields
)
867 log(opcode
, bin(insn_num
))
868 yield ".long 0x%X # %s" % (insn_num
, insn
)
871 # identify if is a svp64 mnemonic
872 if not opcode
.startswith('sv.'):
873 yield insn
# unaltered
875 opcode
= opcode
[3:] # strip leading "sv"
877 # start working on decoding the svp64 op: sv.basev30Bop/vec2/mode
878 opmodes
= opcode
.split("/") # split at "/"
879 v30b_op_orig
= opmodes
.pop(0) # first is the v3.0B
880 # check instruction ends with dot
881 rc_mode
= v30b_op_orig
.endswith('.')
883 v30b_op
= v30b_op_orig
[:-1]
885 v30b_op
= v30b_op_orig
887 # look up the 32-bit op (original, with "." if it has it)
888 if v30b_op_orig
in isa
.instr
:
889 isa_instr
= isa
.instr
[v30b_op_orig
]
891 raise Exception("opcode %s of '%s' not supported" %
892 (v30b_op_orig
, insn
))
894 # look up the svp64 op, first the original (with "." if it has it)
895 if v30b_op_orig
in svp64
.instrs
:
896 rm
= svp64
.instrs
[v30b_op_orig
] # one row of the svp64 RM CSV
897 # then without the "." (if there was one)
898 elif v30b_op
in svp64
.instrs
:
899 rm
= svp64
.instrs
[v30b_op
] # one row of the svp64 RM CSV
901 raise Exception(f
"opcode {v30b_op_orig!r} of "
902 f
"{insn!r} not an svp64 instruction")
904 # get regs info e.g. "RT,RA,RB"
905 v30b_regs
= isa_instr
.regs
[0]
906 log("v3.0B op", v30b_op
, "Rc=1" if rc_mode
else '')
907 log("v3.0B regs", opcode
, v30b_regs
)
910 # right. the first thing to do is identify the ordering of
911 # the registers, by name. the EXTRA2/3 ordering is in
912 # rm['0']..rm['3'] but those fields contain the names RA, BB
913 # etc. we have to read the pseudocode to understand which
914 # reg is which in our instruction. sigh.
916 # first turn the svp64 rm into a "by name" dict, recording
917 # which position in the RM EXTRA it goes into
918 # also: record if the src or dest was a CR, for sanity-checking
919 # (elwidth overrides on CRs are banned)
920 decode
= decode_extra(rm
)
921 dest_reg_cr
, src_reg_cr
, svp64_src
, svp64_dest
= decode
923 log("EXTRA field index, src", svp64_src
)
924 log("EXTRA field index, dest", svp64_dest
)
926 # okaaay now we identify the field value (opcode N,N,N) with
927 # the pseudo-code info (opcode RT, RA, RB)
928 assert len(fields
) == len(v30b_regs
), \
929 "length of fields %s must match insn `%s` fields %s" % \
930 (str(v30b_regs
), insn
, str(fields
))
931 opregfields
= zip(fields
, v30b_regs
) # err that was easy
933 # now for each of those find its place in the EXTRA encoding
934 # note there is the possibility (for LD/ST-with-update) of
935 # RA occurring **TWICE**. to avoid it getting added to the
936 # v3.0B suffix twice, we spot it as a duplicate, here
937 extras
= OrderedDict()
938 for idx
, (field
, regname
) in enumerate(opregfields
):
939 imm
, regname
= decode_imm(regname
)
940 rtype
= get_regtype(regname
)
941 log(" idx find", rtype
, idx
, field
, regname
, imm
)
943 # probably an immediate field, append it straight
944 extras
[('imm', idx
, False)] = (idx
, field
, None, None, None)
946 extra
= svp64_src
.get(regname
, None)
947 if extra
is not None:
948 extra
= ('s', extra
, False) # not a duplicate
949 extras
[extra
] = (idx
, field
, regname
, rtype
, imm
)
950 log(" idx src", idx
, extra
, extras
[extra
])
951 dextra
= svp64_dest
.get(regname
, None)
952 log("regname in", regname
, dextra
)
953 if dextra
is not None:
954 is_a_duplicate
= extra
is not None # duplicate spotted
955 dextra
= ('d', dextra
, is_a_duplicate
)
956 extras
[dextra
] = (idx
, field
, regname
, rtype
, imm
)
957 log(" idx dst", idx
, extra
, extras
[dextra
])
959 # great! got the extra fields in their associated positions:
960 # also we know the register type. now to create the EXTRA encodings
961 etype
= rm
['Etype'] # Extra type: EXTRA3/EXTRA2
962 ptype
= rm
['Ptype'] # Predication type: Twin / Single
965 for extra_idx
, (idx
, field
, rname
, rtype
, iname
) in extras
.items():
966 # is it a field we don't alter/examine? if so just put it
969 v30b_newfields
.append(field
)
972 # identify if this is a ld/st immediate(reg) thing
973 ldst_imm
= "(" in field
and field
[-1] == ')'
975 immed
, field
= field
[:-1].split("(")
977 field
, regmode
= decode_reg(field
, macros
=macros
)
978 log(" ", extra_idx
, rname
, rtype
,
979 regmode
, iname
, field
, end
=" ")
981 # see Mode field https://libre-soc.org/openpower/sv/svp64/
982 # XXX TODO: the following is a bit of a laborious repeated
983 # mess, which could (and should) easily be parameterised.
984 # XXX also TODO: the LD/ST modes which are different
985 # https://libre-soc.org/openpower/sv/ldst/
987 # rright. SVP64 register numbering is from 0 to 127
988 # for GPRs, FPRs *and* CR Fields, where for v3.0 the GPRs and RPFs
989 # are 0-31 and CR Fields are only 0-7. the SVP64 RM "Extra"
990 # area is used to extend the numbering from the 32-bit
991 # instruction, and also to record whether the register
992 # is scalar or vector. on a per-operand basis. this
993 # results in a slightly finnicky encoding: here we go...
995 # encode SV-GPR and SV-FPR field into extra, v3.0field
996 if rtype
in ['GPR', 'FPR']:
997 sv_extra
, field
= get_extra_gpr(etype
, regmode
, field
)
998 # now sanity-check. EXTRA3 is ok, EXTRA2 has limits
999 # (and shrink to a single bit if ok)
1000 if etype
== 'EXTRA2':
1001 if regmode
== 'scalar':
1002 # range is r0-r63 in increments of 1
1003 assert (sv_extra
>> 1) == 0, \
1004 "scalar GPR %s cannot fit into EXTRA2 %s" % \
1005 (rname
, str(extras
[extra_idx
]))
1006 # all good: encode as scalar
1007 sv_extra
= sv_extra
& 0b01
1009 # range is r0-r127 in increments of 2 (r0 r2 ... r126)
1010 assert sv_extra
& 0b01 == 0, \
1011 "%s: vector field %s cannot fit " \
1012 "into EXTRA2 %s" % \
1013 (insn
, rname
, str(extras
[extra_idx
]))
1014 # all good: encode as vector (bit 2 set)
1015 sv_extra
= 0b10 |
(sv_extra
>> 1)
1016 elif regmode
== 'vector':
1017 # EXTRA3 vector bit needs marking
1020 # encode SV-CR 3-bit field into extra, v3.0field.
1021 # 3-bit is for things like BF and BFA
1022 elif rtype
== 'CR_3bit':
1023 sv_extra
, field
= crf_extra(etype
, rname
, extra_idx
,
1024 regmode
, field
, extras
)
1026 # encode SV-CR 5-bit field into extra, v3.0field
1027 # 5-bit is for things like BA BB BC BT etc.
1028 # *sigh* this is the same as 3-bit except the 2 LSBs of the
1029 # 5-bit field are passed through unaltered.
1030 elif rtype
== 'CR_5bit':
1031 cr_subfield
= field
& 0b11 # record bottom 2 bits for later
1032 field
= field
>> 2 # strip bottom 2 bits
1033 # use the exact same 3-bit function for the top 3 bits
1034 sv_extra
, field
= crf_extra(etype
, rname
, extra_idx
,
1035 regmode
, field
, extras
)
1036 # reconstruct the actual 5-bit CR field (preserving the
1037 # bottom 2 bits, unaltered)
1038 field
= (field
<< 2) | cr_subfield
1041 raise Exception("no type match: %s" % rtype
)
1043 # capture the extra field info
1044 log("=>", "%5s" % bin(sv_extra
), field
)
1045 extras
[extra_idx
] = sv_extra
1047 # append altered field value to v3.0b, differs for LDST
1048 # note that duplicates are skipped e.g. EXTRA2 contains
1049 # *BOTH* s:RA *AND* d:RA which happens on LD/ST-with-update
1050 srcdest
, idx
, duplicate
= extra_idx
1051 if duplicate
: # skip adding to v3.0b fields, already added
1054 v30b_newfields
.append(("%s(%s)" % (immed
, str(field
))))
1056 v30b_newfields
.append(str(field
))
1058 log("new v3.0B fields", v30b_op
, v30b_newfields
)
1059 log("extras", extras
)
1061 # rright. now we have all the info. start creating SVP64 instruction.
1062 svp64_insn
= SVP64Instruction
.pair(prefix
=0, suffix
=0)
1063 svp64_prefix
= svp64_insn
.prefix
1064 svp64_rm
= svp64_insn
.prefix
.rm
1066 # begin with EXTRA fields
1067 for idx
, sv_extra
in extras
.items():
1073 srcdest
, idx
, duplicate
= idx
1074 if etype
== 'EXTRA2':
1075 svp64_rm
.extra2
[idx
] = sv_extra
1077 svp64_rm
.extra3
[idx
] = sv_extra
1079 # identify if the op is a LD/ST.
1080 # see https://libre-soc.org/openpower/sv/ldst/
1081 is_ldst
= rm
['mode'] in ['LDST_IDX', 'LDST_IMM']
1082 is_ldst_idx
= rm
['mode'] == 'LDST_IDX'
1083 is_ldst_imm
= rm
['mode'] == 'LDST_IMM'
1084 is_ld
= v30b_op
.startswith("l") and is_ldst
1085 is_st
= v30b_op
.startswith("s") and is_ldst
1087 # branch-conditional or CR detection
1088 is_bc
= rm
['mode'] == 'BRANCH'
1089 is_cr
= rm
['mode'] == 'CROP'
1093 pmask
= 0 # bits 1-3
1094 destwid
= 0 # bits 4-5
1095 srcwid
= 0 # bits 6-7
1096 subvl
= 0 # bits 8-9
1097 smask
= 0 # bits 16-18 but only for twin-predication
1098 mode
= 0 # bits 19-23
1100 mask_m_specified
= False
1110 reverse_gear
= False
1111 mapreduce_crm
= False
1122 # ok let's start identifying opcode augmentation fields
1123 for encmode
in opmodes
:
1124 # predicate mask (src and dest)
1125 if encmode
.startswith("m="):
1127 pmmode
, pmask
= decode_predicate(encmode
[2:])
1128 smmode
, smask
= pmmode
, pmask
1130 mask_m_specified
= True
1131 # predicate mask (dest)
1132 elif encmode
.startswith("dm="):
1134 pmmode
, pmask
= decode_predicate(encmode
[3:])
1137 # predicate mask (src, twin-pred)
1138 elif encmode
.startswith("sm="):
1140 smmode
, smask
= decode_predicate(encmode
[3:])
1144 elif encmode
.startswith("vec"):
1145 subvl
= decode_subvl(encmode
[3:])
1146 # elwidth (both src and dest, like mask)
1147 elif encmode
.startswith("w="):
1148 destwid
= decode_elwidth(encmode
[2:])
1149 srcwid
= decode_elwidth(encmode
[2:])
1151 elif encmode
.startswith("dw="):
1152 destwid
= decode_elwidth(encmode
[3:])
1154 elif encmode
.startswith("sw="):
1155 srcwid
= decode_elwidth(encmode
[3:])
1157 elif encmode
== 'pi':
1159 # in indexed mode, set sv_mode=0b00
1160 assert is_ldst_imm
is True
1162 # element-strided LD/ST
1163 elif encmode
== 'els':
1165 # in indexed mode, set sv_mode=0b01
1169 elif encmode
== 'sats':
1170 assert sv_mode
is None
1173 elif encmode
== 'satu':
1174 assert sv_mode
is None
1178 elif encmode
== 'zz': # TODO, a lot more checking on legality
1179 dst_zero
= 1 # NOT on cr_ops, that's RM[6]
1181 elif encmode
== 'sz':
1183 elif encmode
== 'dz':
1186 elif encmode
.startswith("ff="):
1187 assert sv_mode
is None
1188 if is_cr
: # sigh, CROPs is different
1192 failfirst
= decode_ffirst(encmode
[3:])
1193 assert sea
is False, "cannot use failfirst with signed-address"
1194 # predicate-result, interestingly same as fail-first
1195 elif encmode
.startswith("pr="):
1196 assert sv_mode
is None
1198 predresult
= decode_ffirst(encmode
[3:])
1199 # map-reduce mode, reverse-gear
1200 elif encmode
== 'mrr':
1201 assert sv_mode
is None
1206 elif encmode
== 'mr':
1207 assert sv_mode
is None
1210 elif encmode
== 'crm': # CR on map-reduce
1211 assert sv_mode
is None
1213 mapreduce_crm
= True
1214 elif encmode
== 'vli':
1215 assert failfirst
is not False, "VLi only allowed in failfirst"
1217 elif encmode
== 'sea':
1220 assert failfirst
is False, "cannot use ffirst+signed-address"
1222 if encmode
== 'all':
1223 svp64_rm
.branch
.ALL
= 1
1224 elif encmode
== 'snz':
1225 svp64_rm
.branch
.sz
= 1
1226 svp64_rm
.branch
.SNZ
= 1
1227 elif encmode
== 'sl':
1228 svp64_rm
.branch
.SL
= 1
1229 elif encmode
== 'slu':
1230 svp64_rm
.branch
.SLu
= 1
1231 elif encmode
== 'lru':
1232 svp64_rm
.branch
.LRu
= 1
1233 elif encmode
== 'vs':
1234 svp64_rm
.branch
.VLS
= 1
1235 elif encmode
== 'vsi':
1236 svp64_rm
.branch
.VLS
= 1
1237 svp64_rm
.branch
.vls
.VLi
= 1
1238 elif encmode
== 'vsb':
1239 svp64_rm
.branch
.VLS
= 1
1240 svp64_rm
.branch
.vls
.VSb
= 1
1241 elif encmode
== 'vsbi':
1242 svp64_rm
.branch
.VLS
= 1
1243 svp64_rm
.branch
.vls
.VSb
= 1
1244 svp64_rm
.branch
.vls
.VLi
= 1
1245 elif encmode
== 'ctr':
1246 svp64_rm
.branch
.CTR
= 1
1247 elif encmode
== 'cti':
1248 svp64_rm
.branch
.CTR
= 1
1249 svp64_rm
.branch
.ctr
.CTi
= 1
1251 raise AssertionError("unknown encmode %s" % encmode
)
1253 raise AssertionError("unknown encmode %s" % encmode
)
1255 # post-inc only available on ld-with-update
1257 assert "u" in opcode
, "/pi only available on ld/st-update"
1259 # sanity check if dz/zz used in branch-mode
1260 if is_bc
and dst_zero
:
1261 raise AssertionError("dz/zz not supported in branch, use 'sz'")
1263 # check sea *after* all qualifiers are evaluated
1265 assert sv_mode
in (None, 0b00, 0b01)
1268 # since m=xx takes precedence (overrides) sm=xx and dm=xx,
1269 # treat them as mutually exclusive
1270 if mask_m_specified
:
1271 assert not has_smask
,\
1272 "cannot have both source-mask and predicate mask"
1273 assert not has_pmask
,\
1274 "cannot have both dest-mask and predicate mask"
1275 # since the default is INT predication (ALWAYS), if you
1276 # specify one CR mask, you must specify both, to avoid
1277 # mixing INT and CR reg types
1278 if has_pmask
and pmmode
== 1:
1280 "need explicit source-mask in CR twin predication"
1281 if has_smask
and smmode
== 1:
1283 "need explicit dest-mask in CR twin predication"
1284 # sanity-check that 2Pred mask is same mode
1285 if has_pmask
and has_smask
:
1286 assert smmode
== pmmode
, \
1287 "predicate masks %s and %s must be same reg type" % \
1290 # sanity-check that twin-predication mask only specified in 2P mode
1292 assert not has_smask
, \
1293 "source-mask can only be specified on Twin-predicate ops"
1294 assert not has_pmask
, \
1295 "dest-mask can only be specified on Twin-predicate ops"
1297 # construct the mode field, doing sanity-checking along the way
1299 assert has_smask
or mask_m_specified
, \
1300 "src zeroing requires a source predicate"
1302 assert has_pmask
or mask_m_specified
, \
1303 "dest zeroing requires a dest predicate"
1305 # okaaay, so there are 4 different modes, here, which will be
1306 # partly-merged-in: is_ldst is merged in with "normal", but
1307 # is_bc is so different it's done separately. likewise is_cr
1308 # (when it is done). here are the maps:
1310 # for "normal" arithmetic: https://libre-soc.org/openpower/sv/normal/
1312 | 0-1 | 2 | 3 4 | description |
1313 | --- | --- |---------|-------------------------- |
1314 | 00 | 0 | dz sz | simple mode |
1315 | 00 | 1 | 0 RG | scalar reduce mode (mapreduce) |
1316 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
1317 | 01 | inv | VLi RC1 | Rc=0: ffirst z/nonz |
1318 | 10 | N | dz sz | sat mode: N=0/1 u/s |
1319 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
1320 | 11 | inv | zz RC1 | Rc=0: pred-result z/nonz |
1323 # https://libre-soc.org/openpower/sv/ldst/
1324 # for LD/ST-immediate:
1326 | 0-1 | 2 | 3 4 | description |
1327 | --- | --- |---------|--------------------------- |
1328 | 00 | 0 | zz els | normal mode |
1329 | 00 | 1 | pi lf | post-inc, LD-fault-first |
1330 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
1331 | 01 | inv | els RC1 | Rc=0: ffirst z/nonz |
1332 | 10 | N | zz els | sat mode: N=0/1 u/s |
1333 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
1334 | 11 | inv | els RC1 | Rc=0: pred-result z/nonz |
1337 # for LD/ST-indexed (RA+RB):
1339 | 0-1 | 2 | 3 4 | description |
1340 | --- | --- |---------|----------------------------- |
1341 | 00 | SEA | dz sz | normal mode |
1342 | 01 | SEA | dz sz | strided (scalar only source) |
1343 | 10 | N | dz sz | sat mode: N=0/1 u/s |
1344 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
1345 | 11 | inv | dz RC1 | Rc=0: pred-result z/nonz |
1348 # and leaving out branches and cr_ops for now because they're
1350 """ TODO branches and cr_ops
1354 sv_mode
= int(svp64_rm
.mode
[0, 1])
1356 svp64_rm
.branch
.sz
= 1
1359 ######################################
1360 # "element-strided" mode, ldst_idx
1361 if sv_mode
== 0b01 and is_ldst_idx
:
1362 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
1363 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1364 mode |
= sea
<< SVP64MODE
.SEA
# el-strided
1366 ######################################
1368 elif sv_mode
is None:
1369 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
1370 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1372 # TODO: for now, LD/ST-indexed is ignored.
1373 mode |
= ldst_elstride
<< SVP64MODE
.ELS_NORMAL
# el-strided
1375 # TODO, reduce and subvector mode
1376 # 00 1 dz CRM reduce mode (mapreduce), SUBVL=1
1377 # 00 1 SVM CRM subvector reduce mode, SUBVL>1
1381 ######################################
1382 # ldst-immediate "post" (and "load-fault-first" modes)
1383 elif sv_mode
== 0b00 and ldst_postinc
== 1: # (or ldst_ld_ffirst)
1384 mode |
= (0b1 << SVP64MODE
.LDI_POST
) # sets bit 2
1385 mode |
= (ldst_postinc
<< SVP64MODE
.LDI_PI
) # sets post-inc
1387 ######################################
1389 elif sv_mode
== 0b00:
1390 mode |
= (0b1 << SVP64MODE
.REDUCE
) # sets mapreduce
1391 assert dst_zero
== 0, "dest-zero not allowed in mapreduce mode"
1393 mode |
= (0b1 << SVP64MODE
.RG
) # sets Reverse-gear mode
1395 mode |
= (0b1 << SVP64MODE
.CRM
) # sets CRM mode
1396 assert rc_mode
, "CRM only allowed when Rc=1"
1397 # bit of weird encoding to jam zero-pred or SVM mode in.
1398 # SVM mode can be enabled only when SUBVL=2/3/4 (vec2/3/4)
1400 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1402 ######################################
1404 elif failfirst
is not False and not is_cr
: # sv_mode == 0b01:
1405 assert src_zero
== 0, "dest-zero not allowed in failfirst mode"
1406 if failfirst
== 'RC1':
1407 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
1408 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1409 assert rc_mode
== False, "ffirst RC1 only ok when Rc=0"
1410 elif failfirst
== '~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, "ffirst RC1 only ok when Rc=0"
1416 assert dst_zero
== 0, "dst-zero not allowed in ffirst BO"
1417 assert rc_mode
, "ffirst BO only possible when Rc=1"
1418 mode |
= (failfirst
<< SVP64MODE
.BO_LSB
) # set BO
1420 # (crops is really different)
1421 elif failfirst
is not False and is_cr
:
1422 if failfirst
in ['RC1', '~RC1']:
1423 mode |
= (src_zero
<< SVP64MODE
.SZ
) # predicate src-zeroing
1424 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1425 if failfirst
== '~RC1':
1426 mode |
= (0b1 << SVP64MODE
.INV
) # ... with inversion
1428 assert dst_zero
== src_zero
, "dz must equal sz in ffirst BO"
1429 mode |
= (failfirst
<< SVP64MODE
.BO_LSB
) # set BO
1430 svp64_rm
.cr_op
.zz
= dst_zero
1432 sv_mode |
= 1 # set VLI in LSB of 2-bit mode
1433 #svp64_rm.cr_op.vli = 1
1435 ######################################
1436 # "saturation" modes
1437 elif sv_mode
== 0b10:
1438 mode |
= src_zero
<< SVP64MODE
.SZ
# predicate zeroing
1439 mode |
= dst_zero
<< SVP64MODE
.DZ
# predicate zeroing
1440 mode |
= (saturation
<< SVP64MODE
.N
) # signed/us saturation
1442 ######################################
1443 # "predicate-result" modes. err... code-duplication from ffirst
1444 elif sv_mode
== 0b11:
1445 assert src_zero
== 0, "dest-zero not allowed in predresult mode"
1446 if predresult
== 'RC1':
1447 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
1448 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1449 assert rc_mode
== False, "pr-mode RC1 only ok when Rc=0"
1450 elif predresult
== '~RC1':
1451 mode |
= (0b1 << SVP64MODE
.RC1
) # sets RC1 mode
1452 mode |
= (dst_zero
<< SVP64MODE
.DZ
) # predicate dst-zeroing
1453 mode |
= (0b1 << SVP64MODE
.INV
) # ... with inversion
1454 assert rc_mode
== False, "pr-mode RC1 only ok when Rc=0"
1456 assert dst_zero
== 0, "dst-zero not allowed in pr-mode BO"
1457 assert rc_mode
, "pr-mode BO only possible when Rc=1"
1458 mode |
= (predresult
<< SVP64MODE
.BO_LSB
) # set BO
1460 # whewww.... modes all done :)
1461 # now put into svp64_rm, but respect MSB0 order
1463 mode |
= (0b1 << SVP64MODE
.MOD2_LSB
)
1465 mode |
= (0b1 << SVP64MODE
.MOD2_MSB
)
1468 mode |
= (0b1 << SVP64MODE
.SEA
)
1470 # this is a mess. really look forward to replacing it with Insn DB
1472 svp64_rm
.mode
= mode
# mode: bits 19-23
1473 if vli
and not is_cr
:
1474 svp64_rm
.normal
.ffrc0
.VLi
= 1
1476 # put in predicate masks into svp64_rm
1478 svp64_rm
.smask
= smask
# source pred: bits 16-18
1480 # put in elwidths unless cr
1482 svp64_rm
.ewsrc
= srcwid
# srcwid: bits 6-7
1483 svp64_rm
.elwidth
= destwid
# destwid: bits 4-5
1485 svp64_rm
.mmode
= mmode
# mask mode: bit 0
1486 svp64_rm
.mask
= pmask
# 1-pred: bits 1-3
1487 svp64_rm
.subvl
= subvl
# and subvl: bits 8-9
1489 # nice debug printout. (and now for something completely different)
1490 # https://youtu.be/u0WOIwlXE9g?t=146
1491 svp64_rm_value
= int(svp64_rm
)
1492 log("svp64_rm", hex(svp64_rm_value
), bin(svp64_rm_value
))
1493 log(" mmode 0 :", bin(mmode
))
1494 log(" pmask 1-3 :", bin(pmask
))
1495 log(" dstwid 4-5 :", bin(destwid
))
1496 log(" srcwid 6-7 :", bin(srcwid
))
1497 log(" subvl 8-9 :", bin(subvl
))
1498 log(" mode 19-23:", bin(svp64_rm
.mode
))
1499 offs
= 2 if etype
== 'EXTRA2' else 3 # 2 or 3 bits
1500 for idx
, sv_extra
in extras
.items():
1505 srcdest
, idx
, duplicate
= idx
1506 start
= (10+idx
*offs
)
1507 end
= start
+ offs
-1
1508 log(" extra%d %2d-%2d:" % (idx
, start
, end
),
1511 log(" smask 16-17:", bin(smask
))
1514 # update prefix PO and ID (aka PID)
1515 svp64_prefix
.po
= 0x1
1516 svp64_prefix
.id = 0b11
1518 # fiinally yield the svp64 prefix and the thingy. v3.0b opcode
1519 rc
= '.' if rc_mode
else ''
1520 yield ".long 0x%08x" % int(svp64_prefix
)
1521 log(v30b_op
, v30b_newfields
)
1523 v30b_op_rc
= v30b_op
1524 if not v30b_op
.endswith('.'):
1528 # FIXME(lkcl): should sv.svstep be like svstep?
1529 if v30b_op_rc
in ("svstep", "svstep."):
1530 # compensate for `SVi -= 1` in svstep()
1531 v30b_newfields
[1] = str(int(v30b_newfields
[1]) + 1)
1533 custom_insn_hook
= CUSTOM_INSNS
.get(v30b_op_rc
)
1534 if custom_insn_hook
is not None:
1535 fields
= tuple(map(to_number
, v30b_newfields
))
1536 insn_num
= custom_insn_hook(fields
)
1537 log(opcode
, bin(insn_num
))
1538 yield ".long 0x%X # %s" % (insn_num
, insn
)
1540 # argh, sv.fmadds etc. need to be done manually
1541 elif v30b_op
== 'ffmadds':
1542 opcode
= 59 << (32-6) # bits 0..6 (MSB0)
1543 opcode |
= int(v30b_newfields
[0]) << (32-11) # FRT
1544 opcode |
= int(v30b_newfields
[1]) << (32-16) # FRA
1545 opcode |
= int(v30b_newfields
[2]) << (32-21) # FRB
1546 opcode |
= int(v30b_newfields
[3]) << (32-26) # FRC
1547 opcode |
= 0b00101 << (32-31) # bits 26-30
1549 opcode |
= 1 # Rc, bit 31.
1550 yield ".long 0x%x" % opcode
1551 # argh, sv.fdmadds need to be done manually
1552 elif v30b_op
== 'fdmadds':
1553 opcode
= 59 << (32-6) # bits 0..6 (MSB0)
1554 opcode |
= int(v30b_newfields
[0]) << (32-11) # FRT
1555 opcode |
= int(v30b_newfields
[1]) << (32-16) # FRA
1556 opcode |
= int(v30b_newfields
[2]) << (32-21) # FRB
1557 opcode |
= int(v30b_newfields
[3]) << (32-26) # FRC
1558 opcode |
= 0b11011 << (32-31) # bits 26-30
1560 opcode |
= 1 # Rc, bit 31.
1561 yield ".long 0x%x" % opcode
1562 # argh, sv.ffadds etc. need to be done manually
1563 elif v30b_op
== 'ffadds':
1564 opcode
= 59 << (32-6) # bits 0..6 (MSB0)
1565 opcode |
= int(v30b_newfields
[0]) << (32-11) # FRT
1566 opcode |
= int(v30b_newfields
[1]) << (32-16) # FRA
1567 opcode |
= int(v30b_newfields
[2]) << (32-21) # FRB
1568 opcode |
= 0b1111100000 << (32-31) # bits 21-30
1570 opcode |
= 1 # Rc, bit 31.
1571 yield ".long 0x%x" % opcode
1573 if not v30b_op
.endswith('.'):
1575 yield "%s %s" % (v30b_op
, ", ".join(v30b_newfields
))
1576 for (name
, span
) in svp64_insn
.traverse("SVP64"):
1577 value
= svp64_insn
.storage
[span
]
1578 log(name
, f
"{value.value:0{value.bits}b}", span
)
1579 log("new v3.0B fields", v30b_op
, v30b_newfields
)
1581 def translate(self
, lst
):
1583 yield from self
.translate_one(insn
)
1586 def macro_subst(macros
, txt
):
1588 log("subst", txt
, macros
)
1591 for macro
, value
in macros
.items():
1594 replaced
= txt
.replace(macro
, value
)
1595 log("macro", txt
, "replaced", replaced
, macro
, value
)
1598 toreplace
= '%s.s' % macro
1599 if toreplace
== txt
:
1601 replaced
= txt
.replace(toreplace
, "%s.s" % value
)
1602 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1605 toreplace
= '%s.v' % macro
1606 if toreplace
== txt
:
1608 replaced
= txt
.replace(toreplace
, "%s.v" % value
)
1609 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1612 toreplace
= '*%s' % macro
1613 if toreplace
in txt
:
1615 replaced
= txt
.replace(toreplace
, '*%s' % value
)
1616 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1619 toreplace
= '(%s)' % macro
1620 if toreplace
in txt
:
1622 replaced
= txt
.replace(toreplace
, '(%s)' % value
)
1623 log("macro", txt
, "replaced", replaced
, toreplace
, value
)
1626 log(" processed", txt
)
1634 if not line
[0].isspace():
1642 # get an input file and an output file
1646 outfile
= sys
.stdout
1647 # read the whole lot in advance in case of in-place
1648 lines
= list(infile
.readlines())
1649 elif len(args
) != 2:
1650 print("pysvp64asm [infile | -] [outfile | -]", file=sys
.stderr
)
1656 infile
= open(args
[0], "r")
1657 # read the whole lot in advance in case of in-place overwrite
1658 lines
= list(infile
.readlines())
1661 outfile
= sys
.stdout
1663 outfile
= open(args
[1], "w")
1665 # read the line, look for custom insn, process it
1666 macros
= {} # macros which start ".set"
1669 op
= line
.split("#")[0].strip()
1671 if op
.startswith(".set"):
1672 macro
= op
[4:].split(",")
1673 (macro
, value
) = map(str.strip
, macro
)
1674 macros
[macro
] = value
1675 if not op
.startswith('sv.') and not op
.startswith(tuple(CUSTOM_INSNS
)):
1679 (ws
, line
) = get_ws(line
)
1680 lst
= isa
.translate_one(op
, macros
)
1681 lst
= '; '.join(lst
)
1682 outfile
.write("%s%s # %s\n" % (ws
, lst
, op
))
1685 if __name__
== '__main__':
1686 lst
= ['slw 3, 1, 4',
1689 'sv.cmpi 5, 1, 3, 2',
1691 'sv.isel 64.v, 3, 2, 65.v',
1692 'sv.setb/dm=r3/sm=1<<r3 5, 31',
1693 'sv.setb/m=r3 5, 31',
1694 'sv.setb/vec2 5, 31',
1695 'sv.setb/sw=8/ew=16 5, 31',
1696 'sv.extsw./ff=eq 5, 31',
1697 'sv.extsw./satu/sz/dz/sm=r3/dm=r3 5, 31',
1698 'sv.extsw./pr=eq 5.v, 31',
1699 'sv.add. 5.v, 2.v, 1.v',
1700 'sv.add./m=r3 5.v, 2.v, 1.v',
1703 'sv.stw 5.v, 4(1.v)',
1704 'sv.ld 5.v, 4(1.v)',
1705 'setvl. 2, 3, 4, 0, 1, 1',
1706 'sv.setvl. 2, 3, 4, 0, 1, 1',
1709 "sv.stfsu 0.v, 16(4.v)",
1712 "sv.stfsu/els 0.v, 16(4)",
1715 'sv.add./mr 5.v, 2.v, 1.v',
1717 macros
= {'win2': '50', 'win': '60'}
1719 'sv.addi win2.v, win.v, -1',
1720 'sv.add./mrr 5.v, 2.v, 1.v',
1721 #'sv.lhzsh 5.v, 11(9.v), 15',
1722 #'sv.lwzsh 5.v, 11(9.v), 15',
1723 'sv.ffmadds 6.v, 2.v, 4.v, 6.v',
1726 #'sv.fmadds 0.v, 8.v, 16.v, 4.v',
1727 #'sv.ffadds 0.v, 8.v, 4.v',
1728 'svremap 11, 0, 1, 2, 3, 2, 1',
1729 'svshape 8, 1, 1, 1, 0',
1730 'svshape 8, 1, 1, 1, 1',
1733 #'sv.lfssh 4.v, 11(8.v), 15',
1734 #'sv.lwzsh 4.v, 11(8.v), 15',
1735 #'sv.svstep. 2.v, 4, 0',
1736 #'sv.fcfids. 48.v, 64.v',
1737 'sv.fcoss. 80.v, 0.v',
1738 'sv.fcoss. 20.v, 0.v',
1741 'sv.bc/all 3,12,192',
1742 'sv.bclr/vsbi 3,81.v,192',
1743 'sv.ld 5.v, 4(1.v)',
1744 'sv.svstep. 2.v, 4, 0',
1755 'svindex 0,0,1,0,0,0,0',
1758 'sv.svstep./m=r3 2.v, 4, 0',
1759 'ternlogi 0,0,0,0x5',
1771 'sv.andi. *80, *80, 1',
1772 'sv.ffmadds. 6.v, 2.v, 4.v, 6.v', # incorrectly inserted 32-bit op
1773 'sv.ffmadds 6.v, 2.v, 4.v, 6.v', # correctly converted to .long
1774 'svshape2 8, 1, 31, 7, 1, 1',
1775 'sv.ld 5.v, 4(1.v)',
1776 'sv.stw 5.v, 4(1.v)',
1777 'sv.bc/all 3,12,192',
1781 #"sv.cmp/ff=gt *0,*1,*2,0",
1785 isa
= SVP64Asm(lst
, macros
=macros
)
1786 log("list:\n", "\n\t".join(list(isa
)))
1787 # running svp64.py is designed to test hard-coded lists
1788 # (above) - which strictly speaking should all be unit tests.
1789 # if you need to actually do assembler translation at the
1790 # commandline use "pysvp64asm" - see setup.py
1791 # XXX NO. asm_process()