1 """Power ISA Decoder second stage
3 based on Anton Blanchard microwatt decode2.vhdl
5 Note: OP_TRAP is used for exceptions and interrupts (micro-code style) by
6 over-riding the internal opcode when an exception is needed.
9 from nmigen
import Module
, Elaboratable
, Signal
, Mux
, Const
, Cat
, Repl
, Record
10 from nmigen
.cli
import rtlil
11 from nmutil
.util
import sel
13 from nmutil
.picker
import PriorityPicker
14 from nmutil
.iocontrol
import RecordObject
15 from nmutil
.extend
import exts
17 from openpower
.exceptions
import LDSTException
19 from openpower
.decoder
.power_svp64_prefix
import SVP64PrefixDecoder
20 from openpower
.decoder
.power_svp64_extra
import SVP64CRExtra
, SVP64RegExtra
21 from openpower
.decoder
.power_svp64_rm
import SVP64RMModeDecode
22 from openpower
.decoder
.power_regspec_map
import regspec_decode_read
23 from openpower
.decoder
.power_regspec_map
import regspec_decode_write
24 from openpower
.decoder
.power_decoder
import create_pdecode
25 from openpower
.decoder
.power_enums
import (MicrOp
, CryIn
, Function
,
27 LdstLen
, In1Sel
, In2Sel
, In3Sel
,
28 OutSel
, SPRfull
, SPRreduced
,
30 SVEXTRA
, SVEtype
, SVPtype
)
31 from openpower
.decoder
.decode2execute1
import (Decode2ToExecute1Type
, Data
,
33 from openpower
.sv
.svp64
import SVP64Rec
34 from openpower
.consts
import (MSR
, SPEC
, EXTRA2
, EXTRA3
, SVP64P
, field
,
35 SPEC_SIZE
, SPECb
, SPEC_AUG_SIZE
, SVP64CROffs
)
37 from openpower
.consts
import FastRegsEnum
38 from openpower
.consts
import XERRegsEnum
39 from openpower
.consts
import TT
40 from openpower
.state
import CoreState
41 from openpower
.util
import spr_to_fast
44 def decode_spr_num(spr
):
45 return Cat(spr
[5:10], spr
[0:5])
48 def instr_is_priv(m
, op
, insn
):
49 """determines if the instruction is privileged or not
52 is_priv_insn
= Signal(reset_less
=True)
54 with m
.Case(MicrOp
.OP_ATTN
, MicrOp
.OP_MFMSR
, MicrOp
.OP_MTMSRD
,
55 MicrOp
.OP_MTMSR
, MicrOp
.OP_RFID
):
56 comb
+= is_priv_insn
.eq(1)
57 with m
.Case(MicrOp
.OP_TLBIE
) : comb
+= is_priv_insn
.eq(1)
58 with m
.Case(MicrOp
.OP_MFSPR
, MicrOp
.OP_MTSPR
):
59 with m
.If(insn
[20]): # field XFX.spr[-1] i think
60 comb
+= is_priv_insn
.eq(1)
64 class SPRMap(Elaboratable
):
65 """SPRMap: maps POWER9 SPR numbers to internal enum values, fast and slow
68 def __init__(self
, regreduce_en
):
69 self
.regreduce_en
= regreduce_en
75 self
.spr_i
= Signal(10, reset_less
=True)
76 self
.spr_o
= Data(SPR
, name
="spr_o")
77 self
.fast_o
= Data(3, name
="fast_o")
79 def elaborate(self
, platform
):
85 with m
.Switch(self
.spr_i
):
86 for i
, x
in enumerate(SPR
):
88 m
.d
.comb
+= self
.spr_o
.data
.eq(i
)
89 m
.d
.comb
+= self
.spr_o
.ok
.eq(1)
90 for x
, v
in spr_to_fast
.items():
92 m
.d
.comb
+= self
.fast_o
.data
.eq(v
)
93 m
.d
.comb
+= self
.fast_o
.ok
.eq(1)
97 class DecodeA(Elaboratable
):
98 """DecodeA from instruction
100 decodes register RA, implicit and explicit CSRs
103 def __init__(self
, dec
, regreduce_en
):
104 self
.regreduce_en
= regreduce_en
105 if self
.regreduce_en
:
110 self
.sel_in
= Signal(In1Sel
, reset_less
=True)
111 self
.insn_in
= Signal(32, reset_less
=True)
112 self
.reg_out
= Data(5, name
="reg_a")
113 self
.spr_out
= Data(SPR
, "spr_a")
114 self
.fast_out
= Data(3, "fast_a")
115 self
.sv_nz
= Signal(1)
117 def elaborate(self
, platform
):
122 m
.submodules
.sprmap
= sprmap
= SPRMap(self
.regreduce_en
)
124 # select Register A field, if *full 7 bits* are zero (2 more from SVP64)
125 ra
= Signal(5, reset_less
=True)
126 comb
+= ra
.eq(self
.dec
.RA
)
127 with m
.If((self
.sel_in
== In1Sel
.RA
) |
128 ((self
.sel_in
== In1Sel
.RA_OR_ZERO
) &
129 ((ra
!= Const(0, 5)) |
(self
.sv_nz
!= Const(0, 1))))):
130 comb
+= reg
.data
.eq(ra
)
133 # some Logic/ALU ops have RS as the 3rd arg, but no "RA".
134 # moved it to 1st position (in1_sel)... because
135 rs
= Signal(5, reset_less
=True)
136 comb
+= rs
.eq(self
.dec
.RS
)
137 with m
.If(self
.sel_in
== In1Sel
.RS
):
138 comb
+= reg
.data
.eq(rs
)
141 # decode Fast-SPR based on instruction type
142 with m
.Switch(op
.internal_op
):
144 # BC or BCREG: implicit register (CTR) NOTE: same in DecodeOut
145 with m
.Case(MicrOp
.OP_BC
):
146 with m
.If(~self
.dec
.BO
[2]): # 3.0B p38 BO2=0, use CTR reg
148 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.CTR
)
149 comb
+= self
.fast_out
.ok
.eq(1)
150 with m
.Case(MicrOp
.OP_BCREG
):
151 xo9
= self
.dec
.FormXL
.XO
[9] # 3.0B p38 top bit of XO
152 xo5
= self
.dec
.FormXL
.XO
[5] # 3.0B p38
153 with m
.If(xo9
& ~xo5
):
155 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.CTR
)
156 comb
+= self
.fast_out
.ok
.eq(1)
158 # MFSPR move from SPRs
159 with m
.Case(MicrOp
.OP_MFSPR
):
160 spr
= Signal(10, reset_less
=True)
161 comb
+= spr
.eq(decode_spr_num(self
.dec
.SPR
)) # from XFX
162 comb
+= sprmap
.spr_i
.eq(spr
)
163 comb
+= self
.spr_out
.eq(sprmap
.spr_o
)
164 comb
+= self
.fast_out
.eq(sprmap
.fast_o
)
169 class DecodeAImm(Elaboratable
):
170 """DecodeA immediate from instruction
172 decodes register RA, whether immediate-zero, implicit and
173 explicit CSRs. SVP64 mode requires 2 extra bits
176 def __init__(self
, dec
):
178 self
.sel_in
= Signal(In1Sel
, reset_less
=True)
179 self
.immz_out
= Signal(reset_less
=True)
180 self
.sv_nz
= Signal(1) # EXTRA bits from SVP64
182 def elaborate(self
, platform
):
186 # zero immediate requested
187 ra
= Signal(5, reset_less
=True)
188 comb
+= ra
.eq(self
.dec
.RA
)
189 with m
.If((self
.sel_in
== In1Sel
.RA_OR_ZERO
) &
190 (ra
== Const(0, 5)) &
191 (self
.sv_nz
== Const(0, 1))):
192 comb
+= self
.immz_out
.eq(1)
197 class DecodeB(Elaboratable
):
198 """DecodeB from instruction
200 decodes register RB, different forms of immediate (signed, unsigned),
201 and implicit SPRs. register B is basically "lane 2" into the CompUnits.
202 by industry-standard convention, "lane 2" is where fully-decoded
203 immediates are muxed in.
206 def __init__(self
, dec
):
208 self
.sel_in
= Signal(In2Sel
, reset_less
=True)
209 self
.insn_in
= Signal(32, reset_less
=True)
210 self
.reg_out
= Data(7, "reg_b")
211 self
.reg_isvec
= Signal(1, name
="reg_b_isvec") # TODO: in reg_out
212 self
.fast_out
= Data(3, "fast_b")
214 def elaborate(self
, platform
):
220 # select Register B field
221 with m
.Switch(self
.sel_in
):
222 with m
.Case(In2Sel
.RB
):
223 comb
+= reg
.data
.eq(self
.dec
.RB
)
225 with m
.Case(In2Sel
.RS
):
226 # for M-Form shiftrot
227 comb
+= reg
.data
.eq(self
.dec
.RS
)
230 # decode SPR2 based on instruction type
231 # BCREG implicitly uses LR or TAR for 2nd reg
232 # CTR however is already in fast_spr1 *not* 2.
233 with m
.If(op
.internal_op
== MicrOp
.OP_BCREG
):
234 xo9
= self
.dec
.FormXL
.XO
[9] # 3.0B p38 top bit of XO
235 xo5
= self
.dec
.FormXL
.XO
[5] # 3.0B p38
237 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.LR
)
238 comb
+= self
.fast_out
.ok
.eq(1)
240 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.TAR
)
241 comb
+= self
.fast_out
.ok
.eq(1)
246 class DecodeBImm(Elaboratable
):
247 """DecodeB immediate from instruction
249 def __init__(self
, dec
):
251 self
.sel_in
= Signal(In2Sel
, reset_less
=True)
252 self
.imm_out
= Data(64, "imm_b")
254 def elaborate(self
, platform
):
258 # select Register B Immediate
259 with m
.Switch(self
.sel_in
):
260 with m
.Case(In2Sel
.CONST_UI
): # unsigned
261 comb
+= self
.imm_out
.data
.eq(self
.dec
.UI
)
262 comb
+= self
.imm_out
.ok
.eq(1)
263 with m
.Case(In2Sel
.CONST_SI
): # sign-extended 16-bit
264 si
= Signal(16, reset_less
=True)
265 comb
+= si
.eq(self
.dec
.SI
)
266 comb
+= self
.imm_out
.data
.eq(exts(si
, 16, 64))
267 comb
+= self
.imm_out
.ok
.eq(1)
268 with m
.Case(In2Sel
.CONST_SI_HI
): # sign-extended 16+16=32 bit
269 si_hi
= Signal(32, reset_less
=True)
270 comb
+= si_hi
.eq(self
.dec
.SI
<< 16)
271 comb
+= self
.imm_out
.data
.eq(exts(si_hi
, 32, 64))
272 comb
+= self
.imm_out
.ok
.eq(1)
273 with m
.Case(In2Sel
.CONST_UI_HI
): # unsigned
274 ui
= Signal(16, reset_less
=True)
275 comb
+= ui
.eq(self
.dec
.UI
)
276 comb
+= self
.imm_out
.data
.eq(ui
<< 16)
277 comb
+= self
.imm_out
.ok
.eq(1)
278 with m
.Case(In2Sel
.CONST_LI
): # sign-extend 24+2=26 bit
279 li
= Signal(26, reset_less
=True)
280 comb
+= li
.eq(self
.dec
.LI
<< 2)
281 comb
+= self
.imm_out
.data
.eq(exts(li
, 26, 64))
282 comb
+= self
.imm_out
.ok
.eq(1)
283 with m
.Case(In2Sel
.CONST_BD
): # sign-extend (14+2)=16 bit
284 bd
= Signal(16, reset_less
=True)
285 comb
+= bd
.eq(self
.dec
.BD
<< 2)
286 comb
+= self
.imm_out
.data
.eq(exts(bd
, 16, 64))
287 comb
+= self
.imm_out
.ok
.eq(1)
288 with m
.Case(In2Sel
.CONST_DS
): # sign-extended (14+2=16) bit
289 ds
= Signal(16, reset_less
=True)
290 comb
+= ds
.eq(self
.dec
.DS
<< 2)
291 comb
+= self
.imm_out
.data
.eq(exts(ds
, 16, 64))
292 comb
+= self
.imm_out
.ok
.eq(1)
293 with m
.Case(In2Sel
.CONST_M1
): # signed (-1)
294 comb
+= self
.imm_out
.data
.eq(~
Const(0, 64)) # all 1s
295 comb
+= self
.imm_out
.ok
.eq(1)
296 with m
.Case(In2Sel
.CONST_SH
): # unsigned - for shift
297 comb
+= self
.imm_out
.data
.eq(self
.dec
.sh
)
298 comb
+= self
.imm_out
.ok
.eq(1)
299 with m
.Case(In2Sel
.CONST_SH32
): # unsigned - for shift
300 comb
+= self
.imm_out
.data
.eq(self
.dec
.SH32
)
301 comb
+= self
.imm_out
.ok
.eq(1)
306 class DecodeC(Elaboratable
):
307 """DecodeC from instruction
309 decodes register RC. this is "lane 3" into some CompUnits (not many)
312 def __init__(self
, dec
):
314 self
.sel_in
= Signal(In3Sel
, reset_less
=True)
315 self
.insn_in
= Signal(32, reset_less
=True)
316 self
.reg_out
= Data(5, "reg_c")
318 def elaborate(self
, platform
):
324 # select Register C field
325 with m
.Switch(self
.sel_in
):
326 with m
.Case(In3Sel
.RB
):
327 # for M-Form shiftrot
328 comb
+= reg
.data
.eq(self
.dec
.RB
)
330 with m
.Case(In3Sel
.RS
):
331 comb
+= reg
.data
.eq(self
.dec
.RS
)
337 class DecodeOut(Elaboratable
):
338 """DecodeOut from instruction
340 decodes output register RA, RT or SPR
343 def __init__(self
, dec
, regreduce_en
):
344 self
.regreduce_en
= regreduce_en
345 if self
.regreduce_en
:
350 self
.sel_in
= Signal(OutSel
, reset_less
=True)
351 self
.insn_in
= Signal(32, reset_less
=True)
352 self
.reg_out
= Data(5, "reg_o")
353 self
.spr_out
= Data(SPR
, "spr_o")
354 self
.fast_out
= Data(3, "fast_o")
356 def elaborate(self
, platform
):
359 m
.submodules
.sprmap
= sprmap
= SPRMap(self
.regreduce_en
)
363 # select Register out field
364 with m
.Switch(self
.sel_in
):
365 with m
.Case(OutSel
.RT
):
366 comb
+= reg
.data
.eq(self
.dec
.RT
)
368 with m
.Case(OutSel
.RA
):
369 comb
+= reg
.data
.eq(self
.dec
.RA
)
371 with m
.Case(OutSel
.SPR
):
372 spr
= Signal(10, reset_less
=True)
373 comb
+= spr
.eq(decode_spr_num(self
.dec
.SPR
)) # from XFX
374 # MFSPR move to SPRs - needs mapping
375 with m
.If(op
.internal_op
== MicrOp
.OP_MTSPR
):
376 comb
+= sprmap
.spr_i
.eq(spr
)
377 comb
+= self
.spr_out
.eq(sprmap
.spr_o
)
378 comb
+= self
.fast_out
.eq(sprmap
.fast_o
)
381 with m
.Switch(op
.internal_op
):
383 # BC or BCREG: implicit register (CTR) NOTE: same in DecodeA
384 with m
.Case(MicrOp
.OP_BC
, MicrOp
.OP_BCREG
):
385 with m
.If(~self
.dec
.BO
[2]): # 3.0B p38 BO2=0, use CTR reg
387 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.CTR
)
388 comb
+= self
.fast_out
.ok
.eq(1)
390 # RFID 1st spr (fast)
391 with m
.Case(MicrOp
.OP_RFID
):
392 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.SRR0
) # SRR0
393 comb
+= self
.fast_out
.ok
.eq(1)
398 class DecodeOut2(Elaboratable
):
399 """DecodeOut2 from instruction
401 decodes output registers (2nd one). note that RA is *implicit* below,
402 which now causes problems with SVP64
404 TODO: SVP64 is a little more complex, here. svp64 allows extending
405 by one more destination by having one more EXTRA field. RA-as-src
406 is not the same as RA-as-dest. limited in that it's the same first
407 5 bits (from the v3.0B opcode), but still kinda cool. mostly used
408 for operations that have src-as-dest: mostly this is LD/ST-with-update
409 but there are others.
412 def __init__(self
, dec
):
414 self
.sel_in
= Signal(OutSel
, reset_less
=True)
415 self
.lk
= Signal(reset_less
=True)
416 self
.insn_in
= Signal(32, reset_less
=True)
417 self
.reg_out
= Data(5, "reg_o2")
418 self
.fast_out
= Data(3, "fast_o2")
419 self
.fast_out3
= Data(3, "fast_o3")
421 def elaborate(self
, platform
):
425 #m.submodules.svdec = svdec = SVP64RegExtra()
427 # get the 5-bit reg data before svp64-munging it into 7-bit plus isvec
428 #reg = Signal(5, reset_less=True)
430 if hasattr(self
.dec
.op
, "upd"):
431 # update mode LD/ST uses read-reg A also as an output
432 with m
.If(self
.dec
.op
.upd
== LDSTMode
.update
):
433 comb
+= self
.reg_out
.data
.eq(self
.dec
.RA
)
434 comb
+= self
.reg_out
.ok
.eq(1)
436 # B, BC or BCREG: potential implicit register (LR) output
437 # these give bl, bcl, bclrl, etc.
438 with m
.Switch(op
.internal_op
):
440 # BC* implicit register (LR)
441 with m
.Case(MicrOp
.OP_BC
, MicrOp
.OP_B
, MicrOp
.OP_BCREG
):
442 with m
.If(self
.lk
): # "link" mode
443 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.LR
) # LR
444 comb
+= self
.fast_out
.ok
.eq(1)
446 # RFID 2nd and 3rd spr (fast)
447 with m
.Case(MicrOp
.OP_RFID
):
448 comb
+= self
.fast_out
.data
.eq(FastRegsEnum
.SRR1
) # SRR1
449 comb
+= self
.fast_out
.ok
.eq(1)
450 comb
+= self
.fast_out3
.data
.eq(FastRegsEnum
.SVSRR0
) # SVSRR0
451 comb
+= self
.fast_out3
.ok
.eq(1)
456 class DecodeRC(Elaboratable
):
457 """DecodeRc from instruction
459 decodes Record bit Rc
462 def __init__(self
, dec
):
464 self
.sel_in
= Signal(RC
, reset_less
=True)
465 self
.insn_in
= Signal(32, reset_less
=True)
466 self
.rc_out
= Data(1, "rc")
468 def elaborate(self
, platform
):
472 # select Record bit out field
473 with m
.Switch(self
.sel_in
):
475 comb
+= self
.rc_out
.data
.eq(self
.dec
.Rc
)
476 comb
+= self
.rc_out
.ok
.eq(1)
478 comb
+= self
.rc_out
.data
.eq(1)
479 comb
+= self
.rc_out
.ok
.eq(1)
480 with m
.Case(RC
.NONE
):
481 comb
+= self
.rc_out
.data
.eq(0)
482 comb
+= self
.rc_out
.ok
.eq(1)
487 class DecodeOE(Elaboratable
):
488 """DecodeOE from instruction
490 decodes OE field: uses RC decode detection which might not be good
492 -- For now, use "rc" in the decode table to decide whether oe exists.
493 -- This is not entirely correct architecturally: For mulhd and
494 -- mulhdu, the OE field is reserved. It remains to be seen what an
495 -- actual POWER9 does if we set it on those instructions, for now we
496 -- test that further down when assigning to the multiplier oe input.
499 def __init__(self
, dec
):
501 self
.sel_in
= Signal(RC
, reset_less
=True)
502 self
.insn_in
= Signal(32, reset_less
=True)
503 self
.oe_out
= Data(1, "oe")
505 def elaborate(self
, platform
):
510 with m
.Switch(op
.internal_op
):
512 # mulhw, mulhwu, mulhd, mulhdu - these *ignore* OE
514 # XXX ARGH! ignoring OE causes incompatibility with microwatt
515 # http://lists.libre-soc.org/pipermail/libre-soc-dev/2020-August/000302.html
516 with m
.Case(MicrOp
.OP_MUL_H64
, MicrOp
.OP_MUL_H32
,
517 MicrOp
.OP_EXTS
, MicrOp
.OP_CNTZ
,
518 MicrOp
.OP_SHL
, MicrOp
.OP_SHR
, MicrOp
.OP_RLC
,
519 MicrOp
.OP_LOAD
, MicrOp
.OP_STORE
,
520 MicrOp
.OP_RLCL
, MicrOp
.OP_RLCR
,
524 # all other ops decode OE field
526 # select OE bit out field
527 with m
.Switch(self
.sel_in
):
529 comb
+= self
.oe_out
.data
.eq(self
.dec
.OE
)
530 comb
+= self
.oe_out
.ok
.eq(1)
535 class DecodeCRIn(Elaboratable
):
536 """Decodes input CR from instruction
538 CR indices - insn fields - (not the data *in* the CR) require only 3
539 bits because they refer to CR0-CR7
542 def __init__(self
, dec
):
544 self
.sel_in
= Signal(CRInSel
, reset_less
=True)
545 self
.insn_in
= Signal(32, reset_less
=True)
546 self
.cr_bitfield
= Data(3, "cr_bitfield")
547 self
.cr_bitfield_b
= Data(3, "cr_bitfield_b")
548 self
.cr_bitfield_o
= Data(3, "cr_bitfield_o")
549 self
.whole_reg
= Data(8, "cr_fxm")
550 self
.sv_override
= Signal(2, reset_less
=True) # do not do EXTRA spec
552 def elaborate(self
, platform
):
556 m
.submodules
.ppick
= ppick
= PriorityPicker(8, reverse_i
=True,
559 # zero-initialisation
560 comb
+= self
.cr_bitfield
.ok
.eq(0)
561 comb
+= self
.cr_bitfield_b
.ok
.eq(0)
562 comb
+= self
.cr_bitfield_o
.ok
.eq(0)
563 comb
+= self
.whole_reg
.ok
.eq(0)
564 comb
+= self
.sv_override
.eq(0)
566 # select the relevant CR bitfields
567 with m
.Switch(self
.sel_in
):
568 with m
.Case(CRInSel
.NONE
):
569 pass # No bitfield activated
570 with m
.Case(CRInSel
.CR0
):
571 comb
+= self
.cr_bitfield
.data
.eq(0) # CR0 (MSB0 numbering)
572 comb
+= self
.cr_bitfield
.ok
.eq(1)
573 comb
+= self
.sv_override
.eq(1)
574 with m
.Case(CRInSel
.CR1
):
575 comb
+= self
.cr_bitfield
.data
.eq(1) # CR1 (MSB0 numbering)
576 comb
+= self
.cr_bitfield
.ok
.eq(1)
577 comb
+= self
.sv_override
.eq(2)
578 with m
.Case(CRInSel
.BI
):
579 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.BI
[2:5])
580 comb
+= self
.cr_bitfield
.ok
.eq(1)
581 with m
.Case(CRInSel
.BFA
):
582 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.FormX
.BFA
)
583 comb
+= self
.cr_bitfield
.ok
.eq(1)
584 with m
.Case(CRInSel
.BA_BB
):
585 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.BA
[2:5])
586 comb
+= self
.cr_bitfield
.ok
.eq(1)
587 comb
+= self
.cr_bitfield_b
.data
.eq(self
.dec
.BB
[2:5])
588 comb
+= self
.cr_bitfield_b
.ok
.eq(1)
589 comb
+= self
.cr_bitfield_o
.data
.eq(self
.dec
.BT
[2:5])
590 comb
+= self
.cr_bitfield_o
.ok
.eq(1)
591 with m
.Case(CRInSel
.BC
):
592 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.BC
[2:5])
593 comb
+= self
.cr_bitfield
.ok
.eq(1)
594 with m
.Case(CRInSel
.WHOLE_REG
):
595 comb
+= self
.whole_reg
.ok
.eq(1)
596 move_one
= Signal(reset_less
=True)
597 comb
+= move_one
.eq(self
.insn_in
[20]) # MSB0 bit 11
598 with m
.If((op
.internal_op
== MicrOp
.OP_MFCR
) & move_one
):
599 # must one-hot the FXM field
600 comb
+= ppick
.i
.eq(self
.dec
.FXM
)
601 comb
+= self
.whole_reg
.data
.eq(ppick
.o
)
603 # otherwise use all of it
604 comb
+= self
.whole_reg
.data
.eq(0xff)
609 class DecodeCROut(Elaboratable
):
610 """Decodes input CR from instruction
612 CR indices - insn fields - (not the data *in* the CR) require only 3
613 bits because they refer to CR0-CR7
616 def __init__(self
, dec
):
618 self
.rc_in
= Signal(reset_less
=True)
619 self
.sel_in
= Signal(CROutSel
, reset_less
=True)
620 self
.insn_in
= Signal(32, reset_less
=True)
621 self
.cr_bitfield
= Data(3, "cr_bitfield")
622 self
.whole_reg
= Data(8, "cr_fxm")
623 self
.sv_override
= Signal(2, reset_less
=True) # do not do EXTRA spec
625 def elaborate(self
, platform
):
629 m
.submodules
.ppick
= ppick
= PriorityPicker(8, reverse_i
=True,
632 comb
+= self
.cr_bitfield
.ok
.eq(0)
633 comb
+= self
.whole_reg
.ok
.eq(0)
634 comb
+= self
.sv_override
.eq(0)
636 # please note these MUST match (setting of cr_bitfield.ok) exactly
637 # with write_cr0 below in PowerDecoder2. the reason it's separated
638 # is to avoid having duplicate copies of DecodeCROut in multiple
639 # PowerDecoderSubsets. register decoding should be a one-off in
640 # PowerDecoder2. see https://bugs.libre-soc.org/show_bug.cgi?id=606
642 with m
.Switch(self
.sel_in
):
643 with m
.Case(CROutSel
.NONE
):
644 pass # No bitfield activated
645 with m
.Case(CROutSel
.CR0
):
646 comb
+= self
.cr_bitfield
.data
.eq(0) # CR0 (MSB0 numbering)
647 comb
+= self
.cr_bitfield
.ok
.eq(self
.rc_in
) # only when RC=1
648 comb
+= self
.sv_override
.eq(1)
649 with m
.Case(CROutSel
.CR1
):
650 comb
+= self
.cr_bitfield
.data
.eq(1) # CR1 (MSB0 numbering)
651 comb
+= self
.cr_bitfield
.ok
.eq(self
.rc_in
) # only when RC=1
652 comb
+= self
.sv_override
.eq(2)
653 with m
.Case(CROutSel
.BF
):
654 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.FormX
.BF
)
655 comb
+= self
.cr_bitfield
.ok
.eq(1)
656 with m
.Case(CROutSel
.BT
):
657 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.FormXL
.BT
[2:5])
658 comb
+= self
.cr_bitfield
.ok
.eq(1)
659 with m
.Case(CROutSel
.WHOLE_REG
):
660 comb
+= self
.whole_reg
.ok
.eq(1)
661 move_one
= Signal(reset_less
=True)
662 comb
+= move_one
.eq(self
.insn_in
[20])
663 with m
.If((op
.internal_op
== MicrOp
.OP_MTCRF
)):
665 # must one-hot the FXM field
666 comb
+= ppick
.i
.eq(self
.dec
.FXM
)
667 with m
.If(ppick
.en_o
):
668 comb
+= self
.whole_reg
.data
.eq(ppick
.o
)
670 comb
+= self
.whole_reg
.data
.eq(0b00000001) # CR7
672 comb
+= self
.whole_reg
.data
.eq(self
.dec
.FXM
)
674 # otherwise use all of it
675 comb
+= self
.whole_reg
.data
.eq(0xff)
679 # dictionary of Input Record field names that, if they exist,
680 # will need a corresponding CSV Decoder file column (actually, PowerOp)
681 # to be decoded (this includes the single bit names)
682 record_names
= {'insn_type': 'internal_op',
683 'fn_unit': 'function_unit',
687 'imm_data': 'in2_sel',
688 'invert_in': 'inv_a',
689 'invert_out': 'inv_out',
692 'output_carry': 'cry_out',
693 'input_carry': 'cry_in',
694 'is_32bit': 'is_32b',
697 'data_len': 'ldst_len',
698 'byte_reverse': 'br',
699 'sign_extend': 'sgn_ext',
704 class PowerDecodeSubset(Elaboratable
):
705 """PowerDecodeSubset: dynamic subset decoder
707 only fields actually requested are copied over. hence, "subset" (duh).
709 def __init__(self
, dec
, opkls
=None, fn_name
=None, final
=False, state
=None,
710 svp64_en
=True, regreduce_en
=False):
712 self
.svp64_en
= svp64_en
713 self
.regreduce_en
= regreduce_en
715 self
.sv_rm
= SVP64Rec(name
="dec_svp64") # SVP64 RM field
716 self
.sv_a_nz
= Signal(1)
719 self
.fn_name
= fn_name
721 opkls
= Decode2ToOperand
722 self
.do
= opkls(fn_name
)
723 col_subset
= self
.get_col_subset(self
.do
)
725 # only needed for "main" PowerDecode2
727 self
.e
= Decode2ToExecute1Type(name
=self
.fn_name
, do
=self
.do
,
728 regreduce_en
=regreduce_en
)
730 # create decoder if one not already given
732 dec
= create_pdecode(name
=fn_name
, col_subset
=col_subset
,
733 row_subset
=self
.rowsubsetfn
)
736 # state information needed by the Decoder
738 state
= CoreState("dec2")
741 def get_col_subset(self
, do
):
742 subset
= { 'cr_in', 'cr_out', 'rc_sel'} # needed, non-optional
743 for k
, v
in record_names
.items():
746 print ("get_col_subset", self
.fn_name
, do
.fields
, subset
)
749 def rowsubsetfn(self
, opcode
, row
):
750 """select per-Function-Unit subset of opcodes to be processed
752 normally this just looks at the "unit" column. MMU is different
753 in that it processes specific SPR set/get operations that the SPR
756 return (row
['unit'] == self
.fn_name
or
757 # sigh a dreadful hack: MTSPR and MFSPR need to be processed
758 # by the MMU pipeline so we direct those opcodes to MMU **AND**
759 # SPR pipelines, then selectively weed out the SPRs that should
760 # or should not not go to each pipeline, further down.
761 # really this should be done by modifying the CSV syntax
762 # to support multiple tasks (unit column multiple entries)
763 # see https://bugs.libre-soc.org/show_bug.cgi?id=310
764 (self
.fn_name
== 'MMU' and row
['unit'] == 'SPR' and
765 row
['internal op'] in ['OP_MTSPR', 'OP_MFSPR'])
769 ports
= self
.dec
.ports() + self
.e
.ports()
771 ports
+= self
.sv_rm
.ports()
774 def needs_field(self
, field
, op_field
):
779 return hasattr(do
, field
) and self
.op_get(op_field
) is not None
781 def do_copy(self
, field
, val
, final
=False):
782 if final
or self
.final
:
786 if hasattr(do
, field
) and val
is not None:
787 return getattr(do
, field
).eq(val
)
790 def op_get(self
, op_field
):
791 return getattr(self
.dec
.op
, op_field
, None)
793 def elaborate(self
, platform
):
794 if self
.regreduce_en
:
801 op
, do
= self
.dec
.op
, self
.do
802 msr
, cia
, svstate
= state
.msr
, state
.pc
, state
.svstate
803 # fill in for a normal instruction (not an exception)
804 # copy over if non-exception, non-privileged etc. is detected
806 if self
.fn_name
is None:
809 name
= self
.fn_name
+ "tmp"
810 self
.e_tmp
= Decode2ToExecute1Type(name
=name
, opkls
=self
.opkls
,
811 regreduce_en
=self
.regreduce_en
)
813 # set up submodule decoders
814 m
.submodules
.dec
= self
.dec
815 m
.submodules
.dec_rc
= self
.dec_rc
= dec_rc
= DecodeRC(self
.dec
)
816 m
.submodules
.dec_oe
= dec_oe
= DecodeOE(self
.dec
)
818 # copy instruction through...
819 for i
in [do
.insn
, dec_rc
.insn_in
, dec_oe
.insn_in
, ]:
820 comb
+= i
.eq(self
.dec
.opcode_in
)
822 # ...and subdecoders' input fields
823 comb
+= dec_rc
.sel_in
.eq(op
.rc_sel
)
824 comb
+= dec_oe
.sel_in
.eq(op
.rc_sel
) # XXX should be OE sel
827 comb
+= self
.do_copy("msr", msr
)
828 comb
+= self
.do_copy("cia", cia
)
829 comb
+= self
.do_copy("svstate", svstate
)
831 # set up instruction type
832 # no op: defaults to OP_ILLEGAL
833 internal_op
= self
.op_get("internal_op")
834 comb
+= self
.do_copy("insn_type", internal_op
)
836 # function unit for decoded instruction: requires minor redirect
838 fn
= self
.op_get("function_unit")
839 spr
= Signal(10, reset_less
=True)
840 comb
+= spr
.eq(decode_spr_num(self
.dec
.SPR
)) # from XFX
842 # Microwatt doesn't implement the partition table
843 # instead has PRTBL register (SPR) to point to process table
845 is_mmu_spr
= Signal()
846 comb
+= is_spr_mv
.eq((internal_op
== MicrOp
.OP_MTSPR
) |
847 (internal_op
== MicrOp
.OP_MFSPR
))
848 comb
+= is_mmu_spr
.eq((spr
== SPR
.DSISR
.value
) |
849 (spr
== SPR
.DAR
.value
) |
850 (spr
== SPR
.PRTBL
.value
) |
851 (spr
== SPR
.PIDR
.value
))
852 # MMU must receive MMU SPRs
853 with m
.If(is_spr_mv
& (fn
== Function
.SPR
) & is_mmu_spr
):
854 comb
+= self
.do_copy("fn_unit", Function
.NONE
)
855 comb
+= self
.do_copy("insn_type", MicrOp
.OP_ILLEGAL
)
856 # SPR pipe must *not* receive MMU SPRs
857 with m
.Elif(is_spr_mv
& (fn
== Function
.MMU
) & ~is_mmu_spr
):
858 comb
+= self
.do_copy("fn_unit", Function
.NONE
)
859 comb
+= self
.do_copy("insn_type", MicrOp
.OP_ILLEGAL
)
862 comb
+= self
.do_copy("fn_unit", fn
)
865 if self
.needs_field("zero_a", "in1_sel"):
866 m
.submodules
.dec_ai
= dec_ai
= DecodeAImm(self
.dec
)
867 comb
+= dec_ai
.sv_nz
.eq(self
.sv_a_nz
)
868 comb
+= dec_ai
.sel_in
.eq(op
.in1_sel
)
869 comb
+= self
.do_copy("zero_a", dec_ai
.immz_out
) # RA==0 detected
870 if self
.needs_field("imm_data", "in2_sel"):
871 m
.submodules
.dec_bi
= dec_bi
= DecodeBImm(self
.dec
)
872 comb
+= dec_bi
.sel_in
.eq(op
.in2_sel
)
873 comb
+= self
.do_copy("imm_data", dec_bi
.imm_out
) # imm in RB
876 comb
+= self
.do_copy("rc", dec_rc
.rc_out
)
877 comb
+= self
.do_copy("oe", dec_oe
.oe_out
)
879 # CR in/out - note: these MUST match with what happens in
881 rc_out
= self
.dec_rc
.rc_out
.data
882 with m
.Switch(op
.cr_out
):
883 with m
.Case(CROutSel
.CR0
, CROutSel
.CR1
):
884 comb
+= self
.do_copy("write_cr0", rc_out
) # only when RC=1
885 with m
.Case(CROutSel
.BF
, CROutSel
.BT
):
886 comb
+= self
.do_copy("write_cr0", 1)
888 comb
+= self
.do_copy("input_cr", self
.op_get("cr_in")) # CR in
889 comb
+= self
.do_copy("output_cr", self
.op_get("cr_out")) # CR out
891 # decoded/selected instruction flags
892 comb
+= self
.do_copy("data_len", self
.op_get("ldst_len"))
893 comb
+= self
.do_copy("invert_in", self
.op_get("inv_a"))
894 comb
+= self
.do_copy("invert_out", self
.op_get("inv_out"))
895 comb
+= self
.do_copy("input_carry", self
.op_get("cry_in"))
896 comb
+= self
.do_copy("output_carry", self
.op_get("cry_out"))
897 comb
+= self
.do_copy("is_32bit", self
.op_get("is_32b"))
898 comb
+= self
.do_copy("is_signed", self
.op_get("sgn"))
899 lk
= self
.op_get("lk")
902 comb
+= self
.do_copy("lk", self
.dec
.LK
) # XXX TODO: accessor
904 comb
+= self
.do_copy("byte_reverse", self
.op_get("br"))
905 comb
+= self
.do_copy("sign_extend", self
.op_get("sgn_ext"))
906 comb
+= self
.do_copy("ldst_mode", self
.op_get("upd")) # LD/ST mode
911 class PowerDecode2(PowerDecodeSubset
):
912 """PowerDecode2: the main instruction decoder.
914 whilst PowerDecode is responsible for decoding the actual opcode, this
915 module encapsulates further specialist, sparse information and
916 expansion of fields that is inconvenient to have in the CSV files.
917 for example: the encoding of the immediates, which are detected
918 and expanded out to their full value from an annotated (enum)
921 implicit register usage is also set up, here. for example: OP_BC
922 requires implicitly reading CTR, OP_RFID requires implicitly writing
925 in addition, PowerDecoder2 is responsible for detecting whether
926 instructions are illegal (or privileged) or not, and instead of
927 just leaving at that, *replacing* the instruction to execute with
928 a suitable alternative (trap).
930 LDSTExceptions are done the cycle _after_ they're detected (after
931 they come out of LDSTCompUnit). basically despite the instruction
932 being decoded, the results of the decode are completely ignored
933 and "exception.happened" used to set the "actual" instruction to
934 "OP_TRAP". the LDSTException data structure gets filled in,
935 in the CompTrapOpSubset and that's what it fills in SRR.
937 to make this work, TestIssuer must notice "exception.happened"
938 after the (failed) LD/ST and copies the LDSTException info from
939 the output, into here (PowerDecoder2). without incrementing PC.
942 def __init__(self
, dec
, opkls
=None, fn_name
=None, final
=False,
943 state
=None, svp64_en
=True, regreduce_en
=False):
944 super().__init
__(dec
, opkls
, fn_name
, final
, state
, svp64_en
,
946 self
.ldst_exc
= LDSTException("dec2_exc")
949 self
.cr_out_isvec
= Signal(1, name
="cr_out_isvec")
950 self
.cr_in_isvec
= Signal(1, name
="cr_in_isvec")
951 self
.cr_in_b_isvec
= Signal(1, name
="cr_in_b_isvec")
952 self
.cr_in_o_isvec
= Signal(1, name
="cr_in_o_isvec")
953 self
.in1_isvec
= Signal(1, name
="reg_a_isvec")
954 self
.in2_isvec
= Signal(1, name
="reg_b_isvec")
955 self
.in3_isvec
= Signal(1, name
="reg_c_isvec")
956 self
.o_isvec
= Signal(1, name
="reg_o_isvec")
957 self
.o2_isvec
= Signal(1, name
="reg_o2_isvec")
958 self
.no_in_vec
= Signal(1, name
="no_in_vec") # no inputs vector
959 self
.no_out_vec
= Signal(1, name
="no_out_vec") # no outputs vector
960 self
.loop_continue
= Signal(1, name
="loop_continue")
961 self
.rm_dec
= SVP64RMModeDecode("svp64_rm_dec")
963 self
.no_in_vec
= Const(1, 1)
964 self
.no_out_vec
= Const(1, 1)
965 self
.loop_continue
= Const(0, 1)
967 def get_col_subset(self
, opkls
):
968 subset
= super().get_col_subset(opkls
)
969 subset
.add("asmcode")
970 subset
.add("in1_sel")
971 subset
.add("in2_sel")
972 subset
.add("in3_sel")
973 subset
.add("out_sel")
979 subset
.add("sv_out2")
980 subset
.add("sv_cr_in")
981 subset
.add("sv_cr_out")
982 subset
.add("SV_Etype")
983 subset
.add("SV_Ptype")
985 subset
.add("internal_op")
989 def elaborate(self
, platform
):
990 m
= super().elaborate(platform
)
993 e_out
, op
, do_out
= self
.e
, self
.dec
.op
, self
.e
.do
994 dec_spr
, msr
, cia
, ext_irq
= state
.dec
, state
.msr
, state
.pc
, state
.eint
995 rc_out
= self
.dec_rc
.rc_out
.data
999 # fill in for a normal instruction (not an exception)
1000 # copy over if non-exception, non-privileged etc. is detected
1002 # set up submodule decoders
1003 m
.submodules
.dec_a
= dec_a
= DecodeA(self
.dec
, self
.regreduce_en
)
1004 m
.submodules
.dec_b
= dec_b
= DecodeB(self
.dec
)
1005 m
.submodules
.dec_c
= dec_c
= DecodeC(self
.dec
)
1006 m
.submodules
.dec_o
= dec_o
= DecodeOut(self
.dec
, self
.regreduce_en
)
1007 m
.submodules
.dec_o2
= dec_o2
= DecodeOut2(self
.dec
)
1008 m
.submodules
.dec_cr_in
= self
.dec_cr_in
= DecodeCRIn(self
.dec
)
1009 m
.submodules
.dec_cr_out
= self
.dec_cr_out
= DecodeCROut(self
.dec
)
1010 comb
+= dec_a
.sv_nz
.eq(self
.sv_a_nz
)
1013 # and SVP64 Extra decoders
1014 m
.submodules
.crout_svdec
= crout_svdec
= SVP64CRExtra()
1015 m
.submodules
.crin_svdec
= crin_svdec
= SVP64CRExtra()
1016 m
.submodules
.crin_svdec_b
= crin_svdec_b
= SVP64CRExtra()
1017 m
.submodules
.crin_svdec_o
= crin_svdec_o
= SVP64CRExtra()
1018 m
.submodules
.in1_svdec
= in1_svdec
= SVP64RegExtra()
1019 m
.submodules
.in2_svdec
= in2_svdec
= SVP64RegExtra()
1020 m
.submodules
.in3_svdec
= in3_svdec
= SVP64RegExtra()
1021 m
.submodules
.o_svdec
= o_svdec
= SVP64RegExtra()
1022 m
.submodules
.o2_svdec
= o2_svdec
= SVP64RegExtra()
1024 # debug access to crout_svdec (used in get_pdecode_cr_out)
1025 self
.crout_svdec
= crout_svdec
1027 # and SVP64 RM mode decoder
1028 m
.submodules
.sv_rm_dec
= rm_dec
= self
.rm_dec
1030 # get the 5-bit reg data before svp64-munging it into 7-bit plus isvec
1031 reg
= Signal(5, reset_less
=True)
1033 # copy instruction through...
1034 for i
in [do
.insn
, dec_a
.insn_in
, dec_b
.insn_in
,
1035 self
.dec_cr_in
.insn_in
, self
.dec_cr_out
.insn_in
,
1036 dec_c
.insn_in
, dec_o
.insn_in
, dec_o2
.insn_in
]:
1037 comb
+= i
.eq(self
.dec
.opcode_in
)
1040 comb
+= self
.dec_cr_in
.sel_in
.eq(op
.cr_in
)
1041 comb
+= self
.dec_cr_out
.sel_in
.eq(op
.cr_out
)
1042 comb
+= self
.dec_cr_out
.rc_in
.eq(rc_out
)
1045 comb
+= self
.do_copy("read_cr_whole", self
.dec_cr_in
.whole_reg
)
1046 comb
+= self
.do_copy("write_cr_whole", self
.dec_cr_out
.whole_reg
)
1048 # ...and subdecoders' input fields
1049 comb
+= dec_a
.sel_in
.eq(op
.in1_sel
)
1050 comb
+= dec_b
.sel_in
.eq(op
.in2_sel
)
1051 comb
+= dec_c
.sel_in
.eq(op
.in3_sel
)
1052 comb
+= dec_o
.sel_in
.eq(op
.out_sel
)
1053 comb
+= dec_o2
.sel_in
.eq(op
.out_sel
)
1054 if hasattr(do
, "lk"):
1055 comb
+= dec_o2
.lk
.eq(do
.lk
)
1058 # now do the SVP64 munging. op.SV_Etype and op.sv_in1 comes from
1059 # PowerDecoder which in turn comes from LDST-RM*.csv and RM-*.csv
1060 # which in turn were auto-generated by sv_analysis.py
1061 extra
= self
.sv_rm
.extra
# SVP64 extra bits 10:18
1065 comb
+= crout_svdec
.idx
.eq(op
.sv_cr_out
) # SVP64 CR out
1066 comb
+= self
.cr_out_isvec
.eq(crout_svdec
.isvec
)
1069 # CR in - selection slightly different due to shared CR field sigh
1070 cr_a_idx
= Signal(SVEXTRA
)
1071 cr_b_idx
= Signal(SVEXTRA
)
1073 # these change slightly, when decoding BA/BB. really should have
1074 # their own separate CSV column: sv_cr_in1 and sv_cr_in2, but hey
1075 comb
+= cr_a_idx
.eq(op
.sv_cr_in
)
1076 comb
+= cr_b_idx
.eq(SVEXTRA
.NONE
)
1077 with m
.If(op
.sv_cr_in
== SVEXTRA
.Idx_1_2
.value
):
1078 comb
+= cr_a_idx
.eq(SVEXTRA
.Idx1
)
1079 comb
+= cr_b_idx
.eq(SVEXTRA
.Idx2
)
1081 comb
+= self
.cr_in_isvec
.eq(crin_svdec
.isvec
)
1082 comb
+= self
.cr_in_b_isvec
.eq(crin_svdec_b
.isvec
)
1083 comb
+= self
.cr_in_o_isvec
.eq(crin_svdec_o
.isvec
)
1085 # indices are slightly different, BA/BB mess sorted above
1086 comb
+= crin_svdec
.idx
.eq(cr_a_idx
) # SVP64 CR in A
1087 comb
+= crin_svdec_b
.idx
.eq(cr_b_idx
) # SVP64 CR in B
1088 comb
+= crin_svdec_o
.idx
.eq(op
.sv_cr_out
) # SVP64 CR out
1090 # get SVSTATE srcstep (TODO: elwidth etc.) needed below
1091 srcstep
= Signal
.like(self
.state
.svstate
.srcstep
)
1092 dststep
= Signal
.like(self
.state
.svstate
.dststep
)
1093 comb
+= srcstep
.eq(self
.state
.svstate
.srcstep
)
1094 comb
+= dststep
.eq(self
.state
.svstate
.dststep
)
1096 # registers a, b, c and out and out2 (LD/ST EA)
1097 for to_reg
, fromreg
, svdec
, out
in (
1098 (e
.read_reg1
, dec_a
.reg_out
, in1_svdec
, False),
1099 (e
.read_reg2
, dec_b
.reg_out
, in2_svdec
, False),
1100 (e
.read_reg3
, dec_c
.reg_out
, in3_svdec
, False),
1101 (e
.write_reg
, dec_o
.reg_out
, o_svdec
, True),
1102 (e
.write_ea
, dec_o2
.reg_out
, o2_svdec
, True)):
1103 comb
+= svdec
.extra
.eq(extra
) # EXTRA field of SVP64 RM
1104 comb
+= svdec
.etype
.eq(op
.SV_Etype
) # EXTRA2/3 for this insn
1105 comb
+= svdec
.reg_in
.eq(fromreg
.data
) # 3-bit (CR0/BC/BFA)
1106 comb
+= to_reg
.ok
.eq(fromreg
.ok
)
1107 # detect if Vectorised: add srcstep/dststep if yes.
1108 # to_reg is 7-bits, outs get dststep added, ins get srcstep
1109 with m
.If(svdec
.isvec
):
1110 step
= dststep
if out
else srcstep
1111 comb
+= to_reg
.data
.eq(step
+svdec
.reg_out
)
1113 comb
+= to_reg
.data
.eq(svdec
.reg_out
)
1115 comb
+= in1_svdec
.idx
.eq(op
.sv_in1
) # SVP64 reg #1 (in1_sel)
1116 comb
+= in2_svdec
.idx
.eq(op
.sv_in2
) # SVP64 reg #2 (in2_sel)
1117 comb
+= in3_svdec
.idx
.eq(op
.sv_in3
) # SVP64 reg #3 (in3_sel)
1118 comb
+= o_svdec
.idx
.eq(op
.sv_out
) # SVP64 output (out_sel)
1119 comb
+= o2_svdec
.idx
.eq(op
.sv_out2
) # SVP64 output (implicit)
1120 # XXX TODO - work out where this should come from. the problem is
1121 # that LD-with-update is implied (computed from "is instruction in
1122 # "update mode" rather than specified cleanly as its own CSV column
1124 # output reg-is-vectorised (and when no in/out is vectorised)
1125 comb
+= self
.in1_isvec
.eq(in1_svdec
.isvec
)
1126 comb
+= self
.in2_isvec
.eq(in2_svdec
.isvec
)
1127 comb
+= self
.in3_isvec
.eq(in3_svdec
.isvec
)
1128 comb
+= self
.o_isvec
.eq(o_svdec
.isvec
)
1129 comb
+= self
.o2_isvec
.eq(o2_svdec
.isvec
)
1130 # TODO add SPRs here. must be True when *all* are scalar
1131 l
= map(lambda svdec
: svdec
.isvec
, [in1_svdec
, in2_svdec
, in3_svdec
,
1132 crin_svdec
, crin_svdec_b
, crin_svdec_o
])
1133 comb
+= self
.no_in_vec
.eq(~
Cat(*l
).bool()) # all input scalar
1134 l
= map(lambda svdec
: svdec
.isvec
, [o2_svdec
, o_svdec
, crout_svdec
])
1135 comb
+= self
.no_out_vec
.eq(~
Cat(*l
).bool()) # all output scalar
1136 # now create a general-purpose "test" as to whether looping
1137 # should continue. this doesn't include predication bit-tests
1138 loop
= self
.loop_continue
1139 with m
.Switch(op
.SV_Ptype
):
1140 with m
.Case(SVPtype
.P2
.value
):
1142 # TODO: *and cache-inhibited LD/ST!*
1143 comb
+= loop
.eq(~
(self
.no_in_vec | self
.no_out_vec
))
1144 with m
.Case(SVPtype
.P1
.value
):
1145 # single-predication, test relies on dest only
1146 comb
+= loop
.eq(~self
.no_out_vec
)
1148 # not an SV operation, no looping
1151 # condition registers (CR)
1152 for to_reg
, cr
, name
, svdec
, out
in (
1153 (e
.read_cr1
, self
.dec_cr_in
, "cr_bitfield", crin_svdec
, 0),
1154 (e
.read_cr2
, self
.dec_cr_in
, "cr_bitfield_b", crin_svdec_b
, 0),
1155 (e
.read_cr3
, self
.dec_cr_in
, "cr_bitfield_o", crin_svdec_o
, 0),
1156 (e
.write_cr
, self
.dec_cr_out
, "cr_bitfield", crout_svdec
, 1)):
1157 fromreg
= getattr(cr
, name
)
1158 comb
+= svdec
.extra
.eq(extra
) # EXTRA field of SVP64 RM
1159 comb
+= svdec
.etype
.eq(op
.SV_Etype
) # EXTRA2/3 for this insn
1160 comb
+= svdec
.cr_in
.eq(fromreg
.data
) # 3-bit (CR0/BC/BFA)
1161 with m
.If(svdec
.isvec
):
1162 # check if this is CR0 or CR1: treated differently
1163 # (does not "listen" to EXTRA2/3 spec for a start)
1164 # also: the CRs start from completely different locations
1165 step
= dststep
if out
else srcstep
1166 with m
.If(cr
.sv_override
== 1): # CR0
1167 offs
= SVP64CROffs
.CR0
1168 comb
+= to_reg
.data
.eq(step
+offs
)
1169 with m
.Elif(cr
.sv_override
== 2): # CR1
1170 offs
= SVP64CROffs
.CR1
1171 comb
+= to_reg
.data
.eq(step
+1)
1173 comb
+= to_reg
.data
.eq(step
+svdec
.cr_out
) # 7-bit out
1175 comb
+= to_reg
.data
.eq(svdec
.cr_out
) # 7-bit output
1176 comb
+= to_reg
.ok
.eq(fromreg
.ok
)
1178 # sigh must determine if RA is nonzero (7 bit)
1179 comb
+= self
.sv_a_nz
.eq(e
.read_reg1
.data
!= Const(0, 7))
1181 # connect up to/from read/write GPRs
1182 for to_reg
, fromreg
in ((e
.read_reg1
, dec_a
.reg_out
),
1183 (e
.read_reg2
, dec_b
.reg_out
),
1184 (e
.read_reg3
, dec_c
.reg_out
),
1185 (e
.write_reg
, dec_o
.reg_out
),
1186 (e
.write_ea
, dec_o2
.reg_out
)):
1187 comb
+= to_reg
.data
.eq(fromreg
.data
)
1188 comb
+= to_reg
.ok
.eq(fromreg
.ok
)
1190 # connect up to/from read/write CRs
1191 for to_reg
, cr
, name
in (
1192 (e
.read_cr1
, self
.dec_cr_in
, "cr_bitfield", ),
1193 (e
.read_cr2
, self
.dec_cr_in
, "cr_bitfield_b", ),
1194 (e
.read_cr3
, self
.dec_cr_in
, "cr_bitfield_o", ),
1195 (e
.write_cr
, self
.dec_cr_out
, "cr_bitfield", )):
1196 fromreg
= getattr(cr
, name
)
1197 comb
+= to_reg
.data
.eq(fromreg
.data
)
1198 comb
+= to_reg
.ok
.eq(fromreg
.ok
)
1201 comb
+= e
.read_spr1
.eq(dec_a
.spr_out
)
1202 comb
+= e
.write_spr
.eq(dec_o
.spr_out
)
1204 # Fast regs out including SRR0/1/SVSRR0
1205 comb
+= e
.read_fast1
.eq(dec_a
.fast_out
)
1206 comb
+= e
.read_fast2
.eq(dec_b
.fast_out
)
1207 comb
+= e
.write_fast1
.eq(dec_o
.fast_out
) # SRR0 (OP_RFID)
1208 comb
+= e
.write_fast2
.eq(dec_o2
.fast_out
) # SRR1 (ditto)
1209 comb
+= e
.write_fast3
.eq(dec_o2
.fast_out3
) # SVSRR0 (ditto)
1212 # connect up SVP64 RM Mode decoding
1213 fn
= self
.op_get("function_unit")
1214 comb
+= rm_dec
.fn_in
.eq(fn
) # decode needs to know if LD/ST type
1215 comb
+= rm_dec
.ptype_in
.eq(op
.SV_Ptype
) # Single/Twin predicated
1216 comb
+= rm_dec
.rc_in
.eq(rc_out
) # Rc=1
1217 comb
+= rm_dec
.rm_in
.eq(self
.sv_rm
) # SVP64 RM mode
1219 # sigh this is exactly the sort of thing for which the
1220 # decoder is designed to not need. MTSPR, MFSPR and others need
1221 # access to the XER bits. however setting e.oe is not appropriate
1222 with m
.If(op
.internal_op
== MicrOp
.OP_MFSPR
):
1223 comb
+= e
.xer_in
.eq(0b111) # SO, CA, OV
1224 with m
.If(op
.internal_op
== MicrOp
.OP_CMP
):
1225 comb
+= e
.xer_in
.eq(1<<XERRegsEnum
.SO
) # SO
1226 with m
.If(op
.internal_op
== MicrOp
.OP_MTSPR
):
1227 comb
+= e
.xer_out
.eq(1)
1229 # set the trapaddr to 0x700 for a td/tw/tdi/twi operation
1230 with m
.If(op
.internal_op
== MicrOp
.OP_TRAP
):
1231 # *DO NOT* call self.trap here. that would reset absolutely
1232 # everything including destroying read of RA and RB.
1233 comb
+= self
.do_copy("trapaddr", 0x70) # strip first nibble
1235 ####################
1236 # ok so the instruction's been decoded, blah blah, however
1237 # now we need to determine if it's actually going to go ahead...
1238 # *or* if in fact it's a privileged operation, whether there's
1239 # an external interrupt, etc. etc. this is a simple priority
1240 # if-elif-elif sequence. decrement takes highest priority,
1241 # EINT next highest, privileged operation third.
1243 # check if instruction is privileged
1244 is_priv_insn
= instr_is_priv(m
, op
.internal_op
, e
.do
.insn
)
1246 # different IRQ conditions
1247 ext_irq_ok
= Signal()
1248 dec_irq_ok
= Signal()
1251 ldst_exc
= self
.ldst_exc
1253 comb
+= ext_irq_ok
.eq(ext_irq
& msr
[MSR
.EE
]) # v3.0B p944 (MSR.EE)
1254 comb
+= dec_irq_ok
.eq(dec_spr
[63] & msr
[MSR
.EE
]) # 6.5.11 p1076
1255 comb
+= priv_ok
.eq(is_priv_insn
& msr
[MSR
.PR
])
1256 comb
+= illeg_ok
.eq(op
.internal_op
== MicrOp
.OP_ILLEGAL
)
1258 # LD/ST exceptions. TestIssuer copies the exception info at us
1259 # after a failed LD/ST.
1260 with m
.If(ldst_exc
.happened
):
1261 with m
.If(ldst_exc
.alignment
):
1262 self
.trap(m
, TT
.PRIV
, 0x600)
1263 with m
.Elif(ldst_exc
.instr_fault
):
1264 with m
.If(ldst_exc
.segment_fault
):
1265 self
.trap(m
, TT
.PRIV
, 0x480)
1267 # pass exception info to trap to create SRR1
1268 self
.trap(m
, TT
.MEMEXC
, 0x400, ldst_exc
)
1270 with m
.If(ldst_exc
.segment_fault
):
1271 self
.trap(m
, TT
.PRIV
, 0x380)
1273 self
.trap(m
, TT
.PRIV
, 0x300)
1275 # decrement counter (v3.0B p1099): TODO 32-bit version (MSR.LPCR)
1276 with m
.Elif(dec_irq_ok
):
1277 self
.trap(m
, TT
.DEC
, 0x900) # v3.0B 6.5 p1065
1279 # external interrupt? only if MSR.EE set
1280 with m
.Elif(ext_irq_ok
):
1281 self
.trap(m
, TT
.EINT
, 0x500)
1283 # privileged instruction trap
1284 with m
.Elif(priv_ok
):
1285 self
.trap(m
, TT
.PRIV
, 0x700)
1287 # illegal instruction must redirect to trap. this is done by
1288 # *overwriting* the decoded instruction and starting again.
1289 # (note: the same goes for interrupts and for privileged operations,
1290 # just with different trapaddr and traptype)
1291 with m
.Elif(illeg_ok
):
1292 # illegal instruction trap
1293 self
.trap(m
, TT
.ILLEG
, 0x700)
1295 # no exception, just copy things to the output
1299 ####################
1300 # follow-up after trap/irq to set up SRR0/1
1302 # trap: (note e.insn_type so this includes OP_ILLEGAL) set up fast regs
1303 # Note: OP_SC could actually be modified to just be a trap
1304 with m
.If((do_out
.insn_type
== MicrOp
.OP_TRAP
) |
1305 (do_out
.insn_type
== MicrOp
.OP_SC
)):
1306 # TRAP write fast1 = SRR0
1307 comb
+= e_out
.write_fast1
.data
.eq(FastRegsEnum
.SRR0
) # SRR0
1308 comb
+= e_out
.write_fast1
.ok
.eq(1)
1309 # TRAP write fast2 = SRR1
1310 comb
+= e_out
.write_fast2
.data
.eq(FastRegsEnum
.SRR1
) # SRR1
1311 comb
+= e_out
.write_fast2
.ok
.eq(1)
1312 # TRAP write fast2 = SRR1
1313 comb
+= e_out
.write_fast3
.data
.eq(FastRegsEnum
.SVSRR0
) # SVSRR0
1314 comb
+= e_out
.write_fast3
.ok
.eq(1)
1316 # RFID: needs to read SRR0/1
1317 with m
.If(do_out
.insn_type
== MicrOp
.OP_RFID
):
1318 # TRAP read fast1 = SRR0
1319 comb
+= e_out
.read_fast1
.data
.eq(FastRegsEnum
.SRR0
) # SRR0
1320 comb
+= e_out
.read_fast1
.ok
.eq(1)
1321 # TRAP read fast2 = SRR1
1322 comb
+= e_out
.read_fast2
.data
.eq(FastRegsEnum
.SRR1
) # SRR1
1323 comb
+= e_out
.read_fast2
.ok
.eq(1)
1324 # TRAP read fast2 = SVSRR0
1325 comb
+= e_out
.read_fast3
.data
.eq(FastRegsEnum
.SVSRR0
) # SVSRR0
1326 comb
+= e_out
.read_fast3
.ok
.eq(1)
1328 # annoying simulator bug
1329 if hasattr(e_out
, "asmcode") and hasattr(self
.dec
.op
, "asmcode"):
1330 comb
+= e_out
.asmcode
.eq(self
.dec
.op
.asmcode
)
1334 def trap(self
, m
, traptype
, trapaddr
, ldst_exc
=None):
1335 """trap: this basically "rewrites" the decoded instruction as a trap
1338 op
, e
= self
.dec
.op
, self
.e
1339 comb
+= e
.eq(0) # reset eeeeeverything
1342 comb
+= self
.do_copy("insn", self
.dec
.opcode_in
, True)
1343 comb
+= self
.do_copy("insn_type", MicrOp
.OP_TRAP
, True)
1344 comb
+= self
.do_copy("fn_unit", Function
.TRAP
, True)
1345 comb
+= self
.do_copy("trapaddr", trapaddr
>> 4, True) # bottom 4 bits
1346 comb
+= self
.do_copy("traptype", traptype
, True) # request type
1347 comb
+= self
.do_copy("ldst_exc", ldst_exc
, True) # request type
1348 comb
+= self
.do_copy("msr", self
.state
.msr
, True) # copy of MSR "state"
1349 comb
+= self
.do_copy("cia", self
.state
.pc
, True) # copy of PC "state"
1350 comb
+= self
.do_copy("svstate", self
.state
.svstate
, True) # SVSTATE
1354 def get_rdflags(e
, cu
):
1356 for idx
in range(cu
.n_src
):
1357 regfile
, regname
, _
= cu
.get_in_spec(idx
)
1358 rdflag
, read
= regspec_decode_read(e
, regfile
, regname
)
1360 print("rdflags", rdl
)
1364 if __name__
== '__main__':
1365 pdecode
= create_pdecode()
1366 dec2
= PowerDecode2(pdecode
)
1367 vl
= rtlil
.convert(dec2
, ports
=dec2
.ports() + pdecode
.ports())
1368 with
open("dec2.il", "w") as f
: