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 soc
.regfile
.regfiles
import XERRegs
13 from nmutil
.picker
import PriorityPicker
14 from nmutil
.iocontrol
import RecordObject
15 from nmutil
.extend
import exts
17 from soc
.experiment
.mem_types
import LDSTException
19 from soc
.decoder
.power_regspec_map
import regspec_decode_read
20 from soc
.decoder
.power_regspec_map
import regspec_decode_write
21 from soc
.decoder
.power_decoder
import create_pdecode
22 from soc
.decoder
.power_enums
import (MicrOp
, CryIn
, Function
,
24 LdstLen
, In1Sel
, In2Sel
, In3Sel
,
25 OutSel
, SPR
, RC
, LDSTMode
,
27 from soc
.decoder
.decode2execute1
import (Decode2ToExecute1Type
, Data
,
29 from soc
.sv
.svp64
import SVP64Rec
30 from soc
.consts
import (MSR
, sel
, SPEC
, EXTRA2
, EXTRA3
, SVP64P
, field
,
31 SPEC_SIZE
, SPECb
, SPEC_AUG_SIZE
)
33 from soc
.regfile
.regfiles
import FastRegs
34 from soc
.consts
import TT
35 from soc
.config
.state
import CoreState
36 from soc
.regfile
.util
import spr_to_fast
39 def decode_spr_num(spr
):
40 return Cat(spr
[5:10], spr
[0:5])
43 def instr_is_priv(m
, op
, insn
):
44 """determines if the instruction is privileged or not
47 is_priv_insn
= Signal(reset_less
=True)
49 with m
.Case(MicrOp
.OP_ATTN
, MicrOp
.OP_MFMSR
, MicrOp
.OP_MTMSRD
,
50 MicrOp
.OP_MTMSR
, MicrOp
.OP_RFID
):
51 comb
+= is_priv_insn
.eq(1)
52 with m
.Case(MicrOp
.OP_TLBIE
) : comb
+= is_priv_insn
.eq(1)
53 with m
.Case(MicrOp
.OP_MFSPR
, MicrOp
.OP_MTSPR
):
54 with m
.If(insn
[20]): # field XFX.spr[-1] i think
55 comb
+= is_priv_insn
.eq(1)
59 class SPRMap(Elaboratable
):
60 """SPRMap: maps POWER9 SPR numbers to internal enum values, fast and slow
64 self
.spr_i
= Signal(10, reset_less
=True)
65 self
.spr_o
= Data(SPR
, name
="spr_o")
66 self
.fast_o
= Data(3, name
="fast_o")
68 def elaborate(self
, platform
):
70 with m
.Switch(self
.spr_i
):
71 for i
, x
in enumerate(SPR
):
73 m
.d
.comb
+= self
.spr_o
.data
.eq(i
)
74 m
.d
.comb
+= self
.spr_o
.ok
.eq(1)
75 for x
, v
in spr_to_fast
.items():
77 m
.d
.comb
+= self
.fast_o
.data
.eq(v
)
78 m
.d
.comb
+= self
.fast_o
.ok
.eq(1)
82 class SVP64ExtraSpec(Elaboratable
):
83 """SVP64ExtraSpec - decodes SVP64 Extra specification.
85 selects the required EXTRA2/3 field.
87 see https://libre-soc.org/openpower/sv/svp64/
90 self
.extra
= Signal(9, reset_less
=True)
91 self
.etype
= Signal(SVEtype
, reset_less
=True) # 2 or 3 bits
92 self
.idx
= Signal(SVEXTRA
, reset_less
=True) # which part of extra
93 self
.spec
= Signal(3) # EXTRA spec for the register
95 def elaborate(self
, platform
):
101 # back in the LDSTRM-* and RM-* files generated by sv_analysis.py
102 # we marked every op with an Etype: EXTRA2 or EXTRA3, and also said
103 # which of the 4 (or 3 for EXTRA3) sub-fields of bits 10:18 contain
104 # the register-extension information. extract those now
105 with m
.Switch(self
.etype
):
106 # 2-bit index selection mode
107 with m
.Case(SVEtype
.EXTRA2
):
108 with m
.Switch(self
.idx
):
109 with m
.Case(SVEXTRA
.Idx0
): # 1st 2 bits [0:1]
110 comb
+= spec
[SPEC
.VEC
].eq(extra
[EXTRA2
.IDX0_VEC
])
111 comb
+= spec
[SPEC
.MSB
].eq(extra
[EXTRA2
.IDX0_MSB
])
112 with m
.Case(SVEXTRA
.Idx1
): # 2nd 2 bits [2:3]
113 comb
+= spec
[SPEC
.VEC
].eq(extra
[EXTRA2
.IDX1_VEC
])
114 comb
+= spec
[SPEC
.MSB
].eq(extra
[EXTRA2
.IDX1_MSB
])
115 with m
.Case(SVEXTRA
.Idx2
): # 3rd 2 bits [4:5]
116 comb
+= spec
[SPEC
.VEC
].eq(extra
[EXTRA2
.IDX2_VEC
])
117 comb
+= spec
[SPEC
.MSB
].eq(extra
[EXTRA2
.IDX2_MSB
])
118 with m
.Case(SVEXTRA
.Idx3
): # 4th 2 bits [6:7]
119 comb
+= spec
[SPEC
.VEC
].eq(extra
[EXTRA2
.IDX3_VEC
])
120 comb
+= spec
[SPEC
.MSB
].eq(extra
[EXTRA2
.IDX3_MSB
])
121 # 3-bit index selection mode
122 with m
.Case(SVEtype
.EXTRA3
):
123 with m
.Switch(self
.idx
):
124 with m
.Case(SVEXTRA
.Idx0
): # 1st 3 bits [0:2]
125 comb
+= spec
.eq(sel(extra
, EXTRA3
.IDX0
))
126 with m
.Case(SVEXTRA
.Idx1
): # 2nd 3 bits [3:5]
127 comb
+= spec
.eq(sel(extra
, EXTRA3
.IDX1
))
128 with m
.Case(SVEXTRA
.Idx2
): # 3rd 3 bits [6:8]
129 comb
+= spec
.eq(sel(extra
, EXTRA3
.IDX2
))
130 # cannot fit more than 9 bits so there is no 4th thing
135 class SVP64RegExtra(SVP64ExtraSpec
):
136 """SVP64RegExtra - decodes SVP64 Extra fields to determine reg extension
138 incoming 5-bit GPR/FP is turned into a 7-bit and marked as scalar/vector
139 depending on info in one of the positions in the EXTRA field.
141 designed so that "no change" to the 5-bit register number occurs if
142 SV either does not apply or the relevant EXTRA2/3 field bits are zero.
144 see https://libre-soc.org/openpower/sv/svp64/
147 SVP64ExtraSpec
.__init
__(self
)
148 self
.reg_in
= Signal(5) # incoming reg number (5 bits, RA, RB)
149 self
.reg_out
= Signal(7) # extra-augmented output (7 bits)
150 self
.isvec
= Signal(1) # reg is marked as vector if true
152 def elaborate(self
, platform
):
153 m
= super().elaborate(platform
) # select required EXTRA2/3
156 # first get the spec. if not changed it's "scalar identity behaviour"
157 # which is zero which is ok.
160 # now decode it. bit 0 is "scalar/vector". note that spec could be zero
161 # from above, which (by design) has the effect of "no change", below.
163 # simple: isvec is top bit of spec
164 comb
+= self
.isvec
.eq(spec
[SPEC
.VEC
])
165 # extra bits for register number augmentation
166 spec_aug
= Signal(SPEC_AUG_SIZE
)
167 comb
+= spec_aug
.eq(field(spec
, SPECb
.MSB
, SPECb
.LSB
, SPEC_SIZE
))
169 # decode vector differently from scalar
170 with m
.If(self
.isvec
):
171 # Vector: shifted up, extra in LSBs (RA << 2) | spec[1:2]
172 comb
+= self
.reg_out
.eq(Cat(spec_aug
, self
.reg_in
))
174 # Scalar: not shifted up, extra in MSBs RA | (spec[1:2] << 5)
175 comb
+= self
.reg_out
.eq(Cat(self
.reg_in
, spec_aug
))
180 class SVP64CRExtra(SVP64ExtraSpec
):
181 """SVP64CRExtra - decodes SVP64 Extra fields to determine CR extension
183 incoming 3-bit CR is turned into a 7-bit and marked as scalar/vector
184 depending on info in one of the positions in the EXTRA field.
186 yes, really, 128 CRs. INT is 128, FP is 128, therefore CRs are 128.
188 designed so that "no change" to the 3-bit CR register number occurs if
189 SV either does not apply or the relevant EXTRA2/3 field bits are zero.
191 see https://libre-soc.org/openpower/sv/svp64/appendix
194 SVP64ExtraSpec
.__init
__(self
)
195 self
.cr_in
= Signal(3) # incoming CR number (3 bits, BA[0:2], BFA)
196 self
.cr_out
= Signal(7) # extra-augmented CR output (7 bits)
197 self
.isvec
= Signal(1) # reg is marked as vector if true
199 def elaborate(self
, platform
):
200 m
= super().elaborate(platform
) # select required EXTRA2/3
203 # first get the spec. if not changed it's "scalar identity behaviour"
204 # which is zero which is ok.
207 # now decode it. bit 0 is "scalar/vector". note that spec could be zero
208 # from above, which (by design) has the effect of "no change", below.
210 # simple: isvec is top bit of spec
211 comb
+= self
.isvec
.eq(spec
[SPEC
.VEC
])
212 # extra bits for register number augmentation
213 spec_aug
= Signal(SPEC_AUG_SIZE
)
214 comb
+= spec_aug
.eq(field(spec
, SPECb
.MSB
, SPECb
.LSB
, SPEC_SIZE
))
216 # decode vector differently from scalar, insert bits 1 and 2 accordingly
217 with m
.If(self
.isvec
):
218 # Vector: shifted up, extra in LSBs (CR << 4) | (spec[1:2] << 2)
219 comb
+= self
.cr_out
.eq(Cat(Const(0, 2), spec_aug
, self
.cr_in
))
221 # Scalar: not shifted up, extra in MSBs CR | (spec[1:2] << 3)
222 comb
+= self
.cr_out
.eq(Cat(self
.cr_in
, spec_aug
))
227 class DecodeA(Elaboratable
):
228 """DecodeA from instruction
230 decodes register RA, implicit and explicit CSRs
233 def __init__(self
, dec
):
235 self
.sel_in
= Signal(In1Sel
, reset_less
=True)
236 self
.insn_in
= Signal(32, reset_less
=True)
237 self
.reg_out
= Data(5, name
="reg_a")
238 self
.spr_out
= Data(SPR
, "spr_a")
239 self
.fast_out
= Data(3, "fast_a")
241 def elaborate(self
, platform
):
246 m
.submodules
.sprmap
= sprmap
= SPRMap()
248 # select Register A field
249 ra
= Signal(5, reset_less
=True)
250 comb
+= ra
.eq(self
.dec
.RA
)
251 with m
.If((self
.sel_in
== In1Sel
.RA
) |
252 ((self
.sel_in
== In1Sel
.RA_OR_ZERO
) &
253 (ra
!= Const(0, 5)))):
254 comb
+= reg
.data
.eq(ra
)
257 # some Logic/ALU ops have RS as the 3rd arg, but no "RA".
258 # moved it to 1st position (in1_sel)... because
259 rs
= Signal(5, reset_less
=True)
260 comb
+= rs
.eq(self
.dec
.RS
)
261 with m
.If(self
.sel_in
== In1Sel
.RS
):
262 comb
+= reg
.data
.eq(rs
)
265 # decode Fast-SPR based on instruction type
266 with m
.Switch(op
.internal_op
):
268 # BC or BCREG: implicit register (CTR) NOTE: same in DecodeOut
269 with m
.Case(MicrOp
.OP_BC
):
270 with m
.If(~self
.dec
.BO
[2]): # 3.0B p38 BO2=0, use CTR reg
272 comb
+= self
.fast_out
.data
.eq(FastRegs
.CTR
)
273 comb
+= self
.fast_out
.ok
.eq(1)
274 with m
.Case(MicrOp
.OP_BCREG
):
275 xo9
= self
.dec
.FormXL
.XO
[9] # 3.0B p38 top bit of XO
276 xo5
= self
.dec
.FormXL
.XO
[5] # 3.0B p38
277 with m
.If(xo9
& ~xo5
):
279 comb
+= self
.fast_out
.data
.eq(FastRegs
.CTR
)
280 comb
+= self
.fast_out
.ok
.eq(1)
282 # MFSPR move from SPRs
283 with m
.Case(MicrOp
.OP_MFSPR
):
284 spr
= Signal(10, reset_less
=True)
285 comb
+= spr
.eq(decode_spr_num(self
.dec
.SPR
)) # from XFX
286 comb
+= sprmap
.spr_i
.eq(spr
)
287 comb
+= self
.spr_out
.eq(sprmap
.spr_o
)
288 comb
+= self
.fast_out
.eq(sprmap
.fast_o
)
293 class DecodeAImm(Elaboratable
):
294 """DecodeA immediate from instruction
296 decodes register RA, whether immediate-zero, implicit and
300 def __init__(self
, dec
):
302 self
.sel_in
= Signal(In1Sel
, reset_less
=True)
303 self
.immz_out
= Signal(reset_less
=True)
305 def elaborate(self
, platform
):
309 # zero immediate requested
310 ra
= Signal(5, reset_less
=True)
311 comb
+= ra
.eq(self
.dec
.RA
)
312 with m
.If((self
.sel_in
== In1Sel
.RA_OR_ZERO
) & (ra
== Const(0, 5))):
313 comb
+= self
.immz_out
.eq(1)
318 class DecodeB(Elaboratable
):
319 """DecodeB from instruction
321 decodes register RB, different forms of immediate (signed, unsigned),
322 and implicit SPRs. register B is basically "lane 2" into the CompUnits.
323 by industry-standard convention, "lane 2" is where fully-decoded
324 immediates are muxed in.
327 def __init__(self
, dec
):
329 self
.sel_in
= Signal(In2Sel
, reset_less
=True)
330 self
.insn_in
= Signal(32, reset_less
=True)
331 self
.reg_out
= Data(7, "reg_b")
332 self
.reg_isvec
= Signal(1, name
="reg_b_isvec") # TODO: in reg_out
333 self
.fast_out
= Data(3, "fast_b")
335 def elaborate(self
, platform
):
341 # select Register B field
342 with m
.Switch(self
.sel_in
):
343 with m
.Case(In2Sel
.RB
):
344 comb
+= reg
.data
.eq(self
.dec
.RB
)
346 with m
.Case(In2Sel
.RS
):
347 # for M-Form shiftrot
348 comb
+= reg
.data
.eq(self
.dec
.RS
)
351 # decode SPR2 based on instruction type
352 # BCREG implicitly uses LR or TAR for 2nd reg
353 # CTR however is already in fast_spr1 *not* 2.
354 with m
.If(op
.internal_op
== MicrOp
.OP_BCREG
):
355 xo9
= self
.dec
.FormXL
.XO
[9] # 3.0B p38 top bit of XO
356 xo5
= self
.dec
.FormXL
.XO
[5] # 3.0B p38
358 comb
+= self
.fast_out
.data
.eq(FastRegs
.LR
)
359 comb
+= self
.fast_out
.ok
.eq(1)
361 comb
+= self
.fast_out
.data
.eq(FastRegs
.TAR
)
362 comb
+= self
.fast_out
.ok
.eq(1)
367 class DecodeBImm(Elaboratable
):
368 """DecodeB immediate from instruction
370 def __init__(self
, dec
):
372 self
.sel_in
= Signal(In2Sel
, reset_less
=True)
373 self
.imm_out
= Data(64, "imm_b")
375 def elaborate(self
, platform
):
379 # select Register B Immediate
380 with m
.Switch(self
.sel_in
):
381 with m
.Case(In2Sel
.CONST_UI
): # unsigned
382 comb
+= self
.imm_out
.data
.eq(self
.dec
.UI
)
383 comb
+= self
.imm_out
.ok
.eq(1)
384 with m
.Case(In2Sel
.CONST_SI
): # sign-extended 16-bit
385 si
= Signal(16, reset_less
=True)
386 comb
+= si
.eq(self
.dec
.SI
)
387 comb
+= self
.imm_out
.data
.eq(exts(si
, 16, 64))
388 comb
+= self
.imm_out
.ok
.eq(1)
389 with m
.Case(In2Sel
.CONST_SI_HI
): # sign-extended 16+16=32 bit
390 si_hi
= Signal(32, reset_less
=True)
391 comb
+= si_hi
.eq(self
.dec
.SI
<< 16)
392 comb
+= self
.imm_out
.data
.eq(exts(si_hi
, 32, 64))
393 comb
+= self
.imm_out
.ok
.eq(1)
394 with m
.Case(In2Sel
.CONST_UI_HI
): # unsigned
395 ui
= Signal(16, reset_less
=True)
396 comb
+= ui
.eq(self
.dec
.UI
)
397 comb
+= self
.imm_out
.data
.eq(ui
<< 16)
398 comb
+= self
.imm_out
.ok
.eq(1)
399 with m
.Case(In2Sel
.CONST_LI
): # sign-extend 24+2=26 bit
400 li
= Signal(26, reset_less
=True)
401 comb
+= li
.eq(self
.dec
.LI
<< 2)
402 comb
+= self
.imm_out
.data
.eq(exts(li
, 26, 64))
403 comb
+= self
.imm_out
.ok
.eq(1)
404 with m
.Case(In2Sel
.CONST_BD
): # sign-extend (14+2)=16 bit
405 bd
= Signal(16, reset_less
=True)
406 comb
+= bd
.eq(self
.dec
.BD
<< 2)
407 comb
+= self
.imm_out
.data
.eq(exts(bd
, 16, 64))
408 comb
+= self
.imm_out
.ok
.eq(1)
409 with m
.Case(In2Sel
.CONST_DS
): # sign-extended (14+2=16) bit
410 ds
= Signal(16, reset_less
=True)
411 comb
+= ds
.eq(self
.dec
.DS
<< 2)
412 comb
+= self
.imm_out
.data
.eq(exts(ds
, 16, 64))
413 comb
+= self
.imm_out
.ok
.eq(1)
414 with m
.Case(In2Sel
.CONST_M1
): # signed (-1)
415 comb
+= self
.imm_out
.data
.eq(~
Const(0, 64)) # all 1s
416 comb
+= self
.imm_out
.ok
.eq(1)
417 with m
.Case(In2Sel
.CONST_SH
): # unsigned - for shift
418 comb
+= self
.imm_out
.data
.eq(self
.dec
.sh
)
419 comb
+= self
.imm_out
.ok
.eq(1)
420 with m
.Case(In2Sel
.CONST_SH32
): # unsigned - for shift
421 comb
+= self
.imm_out
.data
.eq(self
.dec
.SH32
)
422 comb
+= self
.imm_out
.ok
.eq(1)
427 class DecodeC(Elaboratable
):
428 """DecodeC from instruction
430 decodes register RC. this is "lane 3" into some CompUnits (not many)
433 def __init__(self
, dec
):
435 self
.sel_in
= Signal(In3Sel
, reset_less
=True)
436 self
.insn_in
= Signal(32, reset_less
=True)
437 self
.reg_out
= Data(5, "reg_c")
439 def elaborate(self
, platform
):
445 # select Register C field
446 with m
.Switch(self
.sel_in
):
447 with m
.Case(In3Sel
.RB
):
448 # for M-Form shiftrot
449 comb
+= reg
.data
.eq(self
.dec
.RB
)
451 with m
.Case(In3Sel
.RS
):
452 comb
+= reg
.data
.eq(self
.dec
.RS
)
458 class DecodeOut(Elaboratable
):
459 """DecodeOut from instruction
461 decodes output register RA, RT or SPR
464 def __init__(self
, dec
):
466 self
.sel_in
= Signal(OutSel
, reset_less
=True)
467 self
.insn_in
= Signal(32, reset_less
=True)
468 self
.reg_out
= Data(5, "reg_o")
469 self
.spr_out
= Data(SPR
, "spr_o")
470 self
.fast_out
= Data(3, "fast_o")
472 def elaborate(self
, platform
):
475 m
.submodules
.sprmap
= sprmap
= SPRMap()
479 # select Register out field
480 with m
.Switch(self
.sel_in
):
481 with m
.Case(OutSel
.RT
):
482 comb
+= reg
.data
.eq(self
.dec
.RT
)
484 with m
.Case(OutSel
.RA
):
485 comb
+= reg
.data
.eq(self
.dec
.RA
)
487 with m
.Case(OutSel
.SPR
):
488 spr
= Signal(10, reset_less
=True)
489 comb
+= spr
.eq(decode_spr_num(self
.dec
.SPR
)) # from XFX
490 # MFSPR move to SPRs - needs mapping
491 with m
.If(op
.internal_op
== MicrOp
.OP_MTSPR
):
492 comb
+= sprmap
.spr_i
.eq(spr
)
493 comb
+= self
.spr_out
.eq(sprmap
.spr_o
)
494 comb
+= self
.fast_out
.eq(sprmap
.fast_o
)
497 with m
.Switch(op
.internal_op
):
499 # BC or BCREG: implicit register (CTR) NOTE: same in DecodeA
500 with m
.Case(MicrOp
.OP_BC
, MicrOp
.OP_BCREG
):
501 with m
.If(~self
.dec
.BO
[2]): # 3.0B p38 BO2=0, use CTR reg
503 comb
+= self
.fast_out
.data
.eq(FastRegs
.CTR
)
504 comb
+= self
.fast_out
.ok
.eq(1)
506 # RFID 1st spr (fast)
507 with m
.Case(MicrOp
.OP_RFID
):
508 comb
+= self
.fast_out
.data
.eq(FastRegs
.SRR0
) # constant: SRR0
509 comb
+= self
.fast_out
.ok
.eq(1)
514 class DecodeOut2(Elaboratable
):
515 """DecodeOut2 from instruction
517 decodes output registers (2nd one). note that RA is *implicit* below,
518 which now causes problems with SVP64
520 TODO: SVP64 is a little more complex, here. svp64 allows extending
521 by one more destination by having one more EXTRA field. RA-as-src
522 is not the same as RA-as-dest. limited in that it's the same first
523 5 bits (from the v3.0B opcode), but still kinda cool. mostly used
524 for operations that have src-as-dest: mostly this is LD/ST-with-update
525 but there are others.
528 def __init__(self
, dec
):
530 self
.sel_in
= Signal(OutSel
, reset_less
=True)
531 self
.lk
= Signal(reset_less
=True)
532 self
.insn_in
= Signal(32, reset_less
=True)
533 self
.reg_out
= Data(5, "reg_o2")
534 self
.fast_out
= Data(3, "fast_o2")
536 def elaborate(self
, platform
):
540 #m.submodules.svdec = svdec = SVP64RegExtra()
542 # get the 5-bit reg data before svp64-munging it into 7-bit plus isvec
543 #reg = Signal(5, reset_less=True)
545 if hasattr(self
.dec
.op
, "upd"):
546 # update mode LD/ST uses read-reg A also as an output
547 with m
.If(self
.dec
.op
.upd
== LDSTMode
.update
):
548 comb
+= self
.reg_out
.data
.eq(self
.dec
.RA
)
549 comb
+= self
.reg_out
.ok
.eq(1)
551 # B, BC or BCREG: potential implicit register (LR) output
552 # these give bl, bcl, bclrl, etc.
553 with m
.Switch(op
.internal_op
):
555 # BC* implicit register (LR)
556 with m
.Case(MicrOp
.OP_BC
, MicrOp
.OP_B
, MicrOp
.OP_BCREG
):
557 with m
.If(self
.lk
): # "link" mode
558 comb
+= self
.fast_out
.data
.eq(FastRegs
.LR
) # constant: LR
559 comb
+= self
.fast_out
.ok
.eq(1)
561 # RFID 2nd spr (fast)
562 with m
.Case(MicrOp
.OP_RFID
):
563 comb
+= self
.fast_out
.data
.eq(FastRegs
.SRR1
) # constant: SRR1
564 comb
+= self
.fast_out
.ok
.eq(1)
569 class DecodeRC(Elaboratable
):
570 """DecodeRc from instruction
572 decodes Record bit Rc
575 def __init__(self
, dec
):
577 self
.sel_in
= Signal(RC
, reset_less
=True)
578 self
.insn_in
= Signal(32, reset_less
=True)
579 self
.rc_out
= Data(1, "rc")
581 def elaborate(self
, platform
):
585 # select Record bit out field
586 with m
.Switch(self
.sel_in
):
588 comb
+= self
.rc_out
.data
.eq(self
.dec
.Rc
)
589 comb
+= self
.rc_out
.ok
.eq(1)
591 comb
+= self
.rc_out
.data
.eq(1)
592 comb
+= self
.rc_out
.ok
.eq(1)
593 with m
.Case(RC
.NONE
):
594 comb
+= self
.rc_out
.data
.eq(0)
595 comb
+= self
.rc_out
.ok
.eq(1)
600 class DecodeOE(Elaboratable
):
601 """DecodeOE from instruction
603 decodes OE field: uses RC decode detection which might not be good
605 -- For now, use "rc" in the decode table to decide whether oe exists.
606 -- This is not entirely correct architecturally: For mulhd and
607 -- mulhdu, the OE field is reserved. It remains to be seen what an
608 -- actual POWER9 does if we set it on those instructions, for now we
609 -- test that further down when assigning to the multiplier oe input.
612 def __init__(self
, dec
):
614 self
.sel_in
= Signal(RC
, reset_less
=True)
615 self
.insn_in
= Signal(32, reset_less
=True)
616 self
.oe_out
= Data(1, "oe")
618 def elaborate(self
, platform
):
623 with m
.Switch(op
.internal_op
):
625 # mulhw, mulhwu, mulhd, mulhdu - these *ignore* OE
627 # XXX ARGH! ignoring OE causes incompatibility with microwatt
628 # http://lists.libre-soc.org/pipermail/libre-soc-dev/2020-August/000302.html
629 with m
.Case(MicrOp
.OP_MUL_H64
, MicrOp
.OP_MUL_H32
,
630 MicrOp
.OP_EXTS
, MicrOp
.OP_CNTZ
,
631 MicrOp
.OP_SHL
, MicrOp
.OP_SHR
, MicrOp
.OP_RLC
,
632 MicrOp
.OP_LOAD
, MicrOp
.OP_STORE
,
633 MicrOp
.OP_RLCL
, MicrOp
.OP_RLCR
,
637 # all other ops decode OE field
639 # select OE bit out field
640 with m
.Switch(self
.sel_in
):
642 comb
+= self
.oe_out
.data
.eq(self
.dec
.OE
)
643 comb
+= self
.oe_out
.ok
.eq(1)
648 class DecodeCRIn(Elaboratable
):
649 """Decodes input CR from instruction
651 CR indices - insn fields - (not the data *in* the CR) require only 3
652 bits because they refer to CR0-CR7
655 def __init__(self
, dec
):
657 self
.sel_in
= Signal(CRInSel
, reset_less
=True)
658 self
.insn_in
= Signal(32, reset_less
=True)
659 self
.cr_bitfield
= Data(3, "cr_bitfield")
660 self
.cr_bitfield_b
= Data(3, "cr_bitfield_b")
661 self
.cr_bitfield_o
= Data(3, "cr_bitfield_o")
662 self
.whole_reg
= Data(8, "cr_fxm")
663 self
.sv_override
= Signal(2, reset_less
=True) # do not do EXTRA spec
665 def elaborate(self
, platform
):
669 m
.submodules
.ppick
= ppick
= PriorityPicker(8, reverse_i
=True,
672 # zero-initialisation
673 comb
+= self
.cr_bitfield
.ok
.eq(0)
674 comb
+= self
.cr_bitfield_b
.ok
.eq(0)
675 comb
+= self
.cr_bitfield_o
.ok
.eq(0)
676 comb
+= self
.whole_reg
.ok
.eq(0)
677 comb
+= self
.sv_override
.eq(0)
679 # select the relevant CR bitfields
680 with m
.Switch(self
.sel_in
):
681 with m
.Case(CRInSel
.NONE
):
682 pass # No bitfield activated
683 with m
.Case(CRInSel
.CR0
):
684 comb
+= self
.cr_bitfield
.data
.eq(0) # CR0 (MSB0 numbering)
685 comb
+= self
.cr_bitfield
.ok
.eq(1)
686 comb
+= self
.sv_override
.eq(1)
687 with m
.Case(CRInSel
.CR1
):
688 comb
+= self
.cr_bitfield
.data
.eq(1) # CR1 (MSB0 numbering)
689 comb
+= self
.cr_bitfield
.ok
.eq(1)
690 comb
+= self
.sv_override
.eq(2)
691 with m
.Case(CRInSel
.BI
):
692 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.BI
[2:5])
693 comb
+= self
.cr_bitfield
.ok
.eq(1)
694 with m
.Case(CRInSel
.BFA
):
695 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.FormX
.BFA
)
696 comb
+= self
.cr_bitfield
.ok
.eq(1)
697 with m
.Case(CRInSel
.BA_BB
):
698 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.BA
[2:5])
699 comb
+= self
.cr_bitfield
.ok
.eq(1)
700 comb
+= self
.cr_bitfield_b
.data
.eq(self
.dec
.BB
[2:5])
701 comb
+= self
.cr_bitfield_b
.ok
.eq(1)
702 comb
+= self
.cr_bitfield_o
.data
.eq(self
.dec
.BT
[2:5])
703 comb
+= self
.cr_bitfield_o
.ok
.eq(1)
704 with m
.Case(CRInSel
.BC
):
705 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.BC
[2:5])
706 comb
+= self
.cr_bitfield
.ok
.eq(1)
707 with m
.Case(CRInSel
.WHOLE_REG
):
708 comb
+= self
.whole_reg
.ok
.eq(1)
709 move_one
= Signal(reset_less
=True)
710 comb
+= move_one
.eq(self
.insn_in
[20]) # MSB0 bit 11
711 with m
.If((op
.internal_op
== MicrOp
.OP_MFCR
) & move_one
):
712 # must one-hot the FXM field
713 comb
+= ppick
.i
.eq(self
.dec
.FXM
)
714 comb
+= self
.whole_reg
.data
.eq(ppick
.o
)
716 # otherwise use all of it
717 comb
+= self
.whole_reg
.data
.eq(0xff)
722 class DecodeCROut(Elaboratable
):
723 """Decodes input CR from instruction
725 CR indices - insn fields - (not the data *in* the CR) require only 3
726 bits because they refer to CR0-CR7
729 def __init__(self
, dec
):
731 self
.rc_in
= Signal(reset_less
=True)
732 self
.sel_in
= Signal(CROutSel
, reset_less
=True)
733 self
.insn_in
= Signal(32, reset_less
=True)
734 self
.cr_bitfield
= Data(3, "cr_bitfield")
735 self
.whole_reg
= Data(8, "cr_fxm")
736 self
.sv_override
= Signal(2, reset_less
=True) # do not do EXTRA spec
738 def elaborate(self
, platform
):
742 m
.submodules
.ppick
= ppick
= PriorityPicker(8, reverse_i
=True,
745 comb
+= self
.cr_bitfield
.ok
.eq(0)
746 comb
+= self
.whole_reg
.ok
.eq(0)
747 comb
+= self
.sv_override
.eq(0)
749 with m
.Switch(self
.sel_in
):
750 with m
.Case(CROutSel
.NONE
):
751 pass # No bitfield activated
752 with m
.Case(CROutSel
.CR0
):
753 comb
+= self
.cr_bitfield
.data
.eq(0) # CR0 (MSB0 numbering)
754 comb
+= self
.cr_bitfield
.ok
.eq(self
.rc_in
) # only when RC=1
755 comb
+= self
.sv_override
.eq(1)
756 with m
.Case(CROutSel
.CR1
):
757 comb
+= self
.cr_bitfield
.data
.eq(1) # CR1 (MSB0 numbering)
758 comb
+= self
.cr_bitfield
.ok
.eq(self
.rc_in
) # only when RC=1
759 comb
+= self
.sv_override
.eq(2)
760 with m
.Case(CROutSel
.BF
):
761 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.FormX
.BF
)
762 comb
+= self
.cr_bitfield
.ok
.eq(1)
763 with m
.Case(CROutSel
.BT
):
764 comb
+= self
.cr_bitfield
.data
.eq(self
.dec
.FormXL
.BT
[2:5])
765 comb
+= self
.cr_bitfield
.ok
.eq(1)
766 with m
.Case(CROutSel
.WHOLE_REG
):
767 comb
+= self
.whole_reg
.ok
.eq(1)
768 move_one
= Signal(reset_less
=True)
769 comb
+= move_one
.eq(self
.insn_in
[20])
770 with m
.If((op
.internal_op
== MicrOp
.OP_MTCRF
)):
772 # must one-hot the FXM field
773 comb
+= ppick
.i
.eq(self
.dec
.FXM
)
774 with m
.If(ppick
.en_o
):
775 comb
+= self
.whole_reg
.data
.eq(ppick
.o
)
777 comb
+= self
.whole_reg
.data
.eq(0b00000001) # CR7
779 comb
+= self
.whole_reg
.data
.eq(self
.dec
.FXM
)
781 # otherwise use all of it
782 comb
+= self
.whole_reg
.data
.eq(0xff)
786 # dictionary of Input Record field names that, if they exist,
787 # will need a corresponding CSV Decoder file column (actually, PowerOp)
788 # to be decoded (this includes the single bit names)
789 record_names
= {'insn_type': 'internal_op',
790 'fn_unit': 'function_unit',
794 'imm_data': 'in2_sel',
795 'invert_in': 'inv_a',
796 'invert_out': 'inv_out',
799 'output_carry': 'cry_out',
800 'input_carry': 'cry_in',
801 'is_32bit': 'is_32b',
804 'data_len': 'ldst_len',
805 'byte_reverse': 'br',
806 'sign_extend': 'sgn_ext',
811 class PowerDecodeSubset(Elaboratable
):
812 """PowerDecodeSubset: dynamic subset decoder
814 only fields actually requested are copied over. hence, "subset" (duh).
816 def __init__(self
, dec
, opkls
=None, fn_name
=None, final
=False, state
=None):
818 self
.sv_rm
= SVP64Rec(name
="dec_svp64") # SVP64 RM field
821 self
.fn_name
= fn_name
823 opkls
= Decode2ToOperand
824 self
.do
= opkls(fn_name
)
825 col_subset
= self
.get_col_subset(self
.do
)
827 # only needed for "main" PowerDecode2
829 self
.e
= Decode2ToExecute1Type(name
=self
.fn_name
, do
=self
.do
)
831 # create decoder if one not already given
833 dec
= create_pdecode(name
=fn_name
, col_subset
=col_subset
,
834 row_subset
=self
.rowsubsetfn
)
837 # state information needed by the Decoder
839 state
= CoreState("dec2")
842 def get_col_subset(self
, do
):
843 subset
= { 'cr_in', 'cr_out', 'rc_sel'} # needed, non-optional
844 for k
, v
in record_names
.items():
847 print ("get_col_subset", self
.fn_name
, do
.fields
, subset
)
850 def rowsubsetfn(self
, opcode
, row
):
851 return row
['unit'] == self
.fn_name
854 return self
.dec
.ports() + self
.e
.ports() + self
.sv_rm
.ports()
856 def needs_field(self
, field
, op_field
):
861 return hasattr(do
, field
) and self
.op_get(op_field
) is not None
863 def do_copy(self
, field
, val
, final
=False):
864 if final
or self
.final
:
868 if hasattr(do
, field
) and val
is not None:
869 return getattr(do
, field
).eq(val
)
872 def op_get(self
, op_field
):
873 return getattr(self
.dec
.op
, op_field
, None)
875 def elaborate(self
, platform
):
879 op
, do
= self
.dec
.op
, self
.do
880 msr
, cia
= state
.msr
, state
.pc
882 # fill in for a normal instruction (not an exception)
883 # copy over if non-exception, non-privileged etc. is detected
885 if self
.fn_name
is None:
888 name
= self
.fn_name
+ "tmp"
889 self
.e_tmp
= Decode2ToExecute1Type(name
=name
, opkls
=self
.opkls
)
891 # set up submodule decoders
892 m
.submodules
.dec
= self
.dec
893 m
.submodules
.dec_rc
= dec_rc
= DecodeRC(self
.dec
)
894 m
.submodules
.dec_oe
= dec_oe
= DecodeOE(self
.dec
)
895 m
.submodules
.dec_cr_in
= self
.dec_cr_in
= DecodeCRIn(self
.dec
)
896 m
.submodules
.dec_cr_out
= self
.dec_cr_out
= DecodeCROut(self
.dec
)
898 # copy instruction through...
900 dec_rc
.insn_in
, dec_oe
.insn_in
,
901 self
.dec_cr_in
.insn_in
, self
.dec_cr_out
.insn_in
]:
902 comb
+= i
.eq(self
.dec
.opcode_in
)
904 # ...and subdecoders' input fields
905 comb
+= dec_rc
.sel_in
.eq(op
.rc_sel
)
906 comb
+= dec_oe
.sel_in
.eq(op
.rc_sel
) # XXX should be OE sel
907 comb
+= self
.dec_cr_in
.sel_in
.eq(op
.cr_in
)
908 comb
+= self
.dec_cr_out
.sel_in
.eq(op
.cr_out
)
909 comb
+= self
.dec_cr_out
.rc_in
.eq(dec_rc
.rc_out
.data
)
912 comb
+= self
.do_copy("msr", msr
)
913 comb
+= self
.do_copy("cia", cia
)
915 # set up instruction type
916 # no op: defaults to OP_ILLEGAL
917 if self
.fn_name
=="MMU":
918 # mmu is special case: needs SPR opcode as well
919 mmu0
= self
.mmu0_spr_dec
920 with m
.If(((mmu0
.dec
.op
.internal_op
== MicrOp
.OP_MTSPR
) |
921 (mmu0
.dec
.op
.internal_op
== MicrOp
.OP_MFSPR
))):
922 comb
+= self
.do_copy("insn_type", mmu0
.op_get("internal_op"))
924 comb
+= self
.do_copy("insn_type", self
.op_get("internal_op"))
926 comb
+= self
.do_copy("insn_type", self
.op_get("internal_op"))
928 # function unit for decoded instruction: requires minor redirect
930 fn
= self
.op_get("function_unit")
931 spr
= Signal(10, reset_less
=True)
932 comb
+= spr
.eq(decode_spr_num(self
.dec
.SPR
)) # from XFX
934 # XXX BUG - don't use hardcoded magic constants.
935 # also use ".value" otherwise the test fails. bit of a pain
936 # https://bugs.libre-soc.org/show_bug.cgi?id=603
938 SPR_PID
= 48 # TODO read docs for POWER9
939 # Microwatt doesn't implement the partition table
940 # instead has PRTBL register (SPR) to point to process table
941 SPR_PRTBL
= 720 # see common.vhdl in microwatt, not in POWER9
942 with m
.If(((self
.dec
.op
.internal_op
== MicrOp
.OP_MTSPR
) |
943 (self
.dec
.op
.internal_op
== MicrOp
.OP_MFSPR
)) &
944 ((spr
== SPR
.DSISR
) |
(spr
== SPR
.DAR
)
945 |
(spr
==SPR_PRTBL
) |
(spr
==SPR_PID
))):
946 comb
+= self
.do_copy("fn_unit", Function
.MMU
)
948 comb
+= self
.do_copy("fn_unit",fn
)
951 if self
.needs_field("zero_a", "in1_sel"):
952 m
.submodules
.dec_ai
= dec_ai
= DecodeAImm(self
.dec
)
953 comb
+= dec_ai
.sel_in
.eq(op
.in1_sel
)
954 comb
+= self
.do_copy("zero_a", dec_ai
.immz_out
) # RA==0 detected
955 if self
.needs_field("imm_data", "in2_sel"):
956 m
.submodules
.dec_bi
= dec_bi
= DecodeBImm(self
.dec
)
957 comb
+= dec_bi
.sel_in
.eq(op
.in2_sel
)
958 comb
+= self
.do_copy("imm_data", dec_bi
.imm_out
) # imm in RB
961 comb
+= self
.do_copy("rc", dec_rc
.rc_out
)
962 comb
+= self
.do_copy("oe", dec_oe
.oe_out
)
965 comb
+= self
.do_copy("read_cr_whole", self
.dec_cr_in
.whole_reg
)
966 comb
+= self
.do_copy("write_cr_whole", self
.dec_cr_out
.whole_reg
)
967 comb
+= self
.do_copy("write_cr0", self
.dec_cr_out
.cr_bitfield
.ok
)
969 comb
+= self
.do_copy("input_cr", self
.op_get("cr_in")) # CR in
970 comb
+= self
.do_copy("output_cr", self
.op_get("cr_out")) # CR out
972 # decoded/selected instruction flags
973 comb
+= self
.do_copy("data_len", self
.op_get("ldst_len"))
974 comb
+= self
.do_copy("invert_in", self
.op_get("inv_a"))
975 comb
+= self
.do_copy("invert_out", self
.op_get("inv_out"))
976 comb
+= self
.do_copy("input_carry", self
.op_get("cry_in"))
977 comb
+= self
.do_copy("output_carry", self
.op_get("cry_out"))
978 comb
+= self
.do_copy("is_32bit", self
.op_get("is_32b"))
979 comb
+= self
.do_copy("is_signed", self
.op_get("sgn"))
980 lk
= self
.op_get("lk")
983 comb
+= self
.do_copy("lk", self
.dec
.LK
) # XXX TODO: accessor
985 comb
+= self
.do_copy("byte_reverse", self
.op_get("br"))
986 comb
+= self
.do_copy("sign_extend", self
.op_get("sgn_ext"))
987 comb
+= self
.do_copy("ldst_mode", self
.op_get("upd")) # LD/ST mode
992 class PowerDecode2(PowerDecodeSubset
):
993 """PowerDecode2: the main instruction decoder.
995 whilst PowerDecode is responsible for decoding the actual opcode, this
996 module encapsulates further specialist, sparse information and
997 expansion of fields that is inconvenient to have in the CSV files.
998 for example: the encoding of the immediates, which are detected
999 and expanded out to their full value from an annotated (enum)
1002 implicit register usage is also set up, here. for example: OP_BC
1003 requires implicitly reading CTR, OP_RFID requires implicitly writing
1006 in addition, PowerDecoder2 is responsible for detecting whether
1007 instructions are illegal (or privileged) or not, and instead of
1008 just leaving at that, *replacing* the instruction to execute with
1009 a suitable alternative (trap).
1011 LDSTExceptions are done the cycle _after_ they're detected (after
1012 they come out of LDSTCompUnit). basically despite the instruction
1013 being decoded, the results of the decode are completely ignored
1014 and "exception.happened" used to set the "actual" instruction to
1015 "OP_TRAP". the LDSTException data structure gets filled in,
1016 in the CompTrapOpSubset and that's what it fills in SRR.
1018 to make this work, TestIssuer must notice "exception.happened"
1019 after the (failed) LD/ST and copies the LDSTException info from
1020 the output, into here (PowerDecoder2). without incrementing PC.
1023 def __init__(self
, dec
, opkls
=None, fn_name
=None, final
=False, state
=None):
1024 super().__init
__(dec
, opkls
, fn_name
, final
, state
)
1025 self
.exc
= LDSTException("dec2_exc")
1027 self
.cr_out_isvec
= Signal(1, name
="cr_out_isvec")
1028 self
.cr_in_isvec
= Signal(1, name
="cr_in_isvec")
1029 self
.cr_in_b_isvec
= Signal(1, name
="cr_in_b_isvec")
1030 self
.cr_in_o_isvec
= Signal(1, name
="cr_in_o_isvec")
1031 self
.in1_isvec
= Signal(1, name
="reg_a_isvec")
1032 self
.in2_isvec
= Signal(1, name
="reg_b_isvec")
1033 self
.in3_isvec
= Signal(1, name
="reg_c_isvec")
1034 self
.o_isvec
= Signal(1, name
="reg_o_isvec")
1035 self
.o2_isvec
= Signal(1, name
="reg_o2_isvec")
1036 self
.no_out_vec
= Signal(1, name
="no_out_vec") # no outputs are vectors
1038 def get_col_subset(self
, opkls
):
1039 subset
= super().get_col_subset(opkls
)
1040 subset
.add("asmcode")
1041 subset
.add("in1_sel")
1042 subset
.add("in2_sel")
1043 subset
.add("in3_sel")
1044 subset
.add("out_sel")
1045 subset
.add("sv_in1")
1046 subset
.add("sv_in2")
1047 subset
.add("sv_in3")
1048 subset
.add("sv_out")
1049 subset
.add("sv_cr_in")
1050 subset
.add("sv_cr_out")
1051 subset
.add("SV_Etype")
1052 subset
.add("SV_Ptype")
1054 subset
.add("internal_op")
1058 def elaborate(self
, platform
):
1059 m
= super().elaborate(platform
)
1062 e_out
, op
, do_out
= self
.e
, self
.dec
.op
, self
.e
.do
1063 dec_spr
, msr
, cia
, ext_irq
= state
.dec
, state
.msr
, state
.pc
, state
.eint
1067 # fill in for a normal instruction (not an exception)
1068 # copy over if non-exception, non-privileged etc. is detected
1070 # set up submodule decoders
1071 m
.submodules
.dec_a
= dec_a
= DecodeA(self
.dec
)
1072 m
.submodules
.dec_b
= dec_b
= DecodeB(self
.dec
)
1073 m
.submodules
.dec_c
= dec_c
= DecodeC(self
.dec
)
1074 m
.submodules
.dec_o
= dec_o
= DecodeOut(self
.dec
)
1075 m
.submodules
.dec_o2
= dec_o2
= DecodeOut2(self
.dec
)
1077 # and SVP64 Extra decoders
1078 m
.submodules
.crout_svdec
= crout_svdec
= SVP64CRExtra()
1079 m
.submodules
.crin_svdec
= crin_svdec
= SVP64CRExtra()
1080 m
.submodules
.crin_svdec_b
= crin_svdec_b
= SVP64CRExtra()
1081 m
.submodules
.crin_svdec_o
= crin_svdec_o
= SVP64CRExtra()
1082 m
.submodules
.in1_svdec
= in1_svdec
= SVP64RegExtra()
1083 m
.submodules
.in2_svdec
= in2_svdec
= SVP64RegExtra()
1084 m
.submodules
.in3_svdec
= in3_svdec
= SVP64RegExtra()
1085 m
.submodules
.o_svdec
= o_svdec
= SVP64RegExtra()
1086 m
.submodules
.o2_svdec
= o2_svdec
= SVP64RegExtra()
1088 # debug access to crout_svdec (used in get_pdecode_cr_out)
1089 self
.crout_svdec
= crout_svdec
1091 # get the 5-bit reg data before svp64-munging it into 7-bit plus isvec
1092 reg
= Signal(5, reset_less
=True)
1094 # copy instruction through...
1095 for i
in [do
.insn
, dec_a
.insn_in
, dec_b
.insn_in
,
1096 dec_c
.insn_in
, dec_o
.insn_in
, dec_o2
.insn_in
]:
1097 comb
+= i
.eq(self
.dec
.opcode_in
)
1099 # now do the SVP64 munging. op.SV_Etype and op.sv_in1 comes from
1100 # PowerDecoder which in turn comes from LDST-RM*.csv and RM-*.csv
1101 # which in turn were auto-generated by sv_analysis.py
1102 extra
= self
.sv_rm
.extra
# SVP64 extra bits 10:18
1106 comb
+= crout_svdec
.idx
.eq(op
.sv_cr_out
) # SVP64 CR out
1107 comb
+= self
.cr_out_isvec
.eq(crout_svdec
.isvec
)
1110 # CR in - index selection slightly different due to shared CR field sigh
1111 cr_a_idx
= Signal(SVEXTRA
)
1112 cr_b_idx
= Signal(SVEXTRA
)
1114 # these change slightly, when decoding BA/BB. really should have
1115 # their own separate CSV column: sv_cr_in1 and sv_cr_in2, but hey
1116 comb
+= cr_a_idx
.eq(op
.sv_cr_in
)
1117 comb
+= cr_b_idx
.eq(SVEXTRA
.NONE
)
1118 with m
.If(op
.sv_cr_in
== SVEXTRA
.Idx_1_2
.value
):
1119 comb
+= cr_a_idx
.eq(SVEXTRA
.Idx1
)
1120 comb
+= cr_b_idx
.eq(SVEXTRA
.Idx2
)
1122 comb
+= self
.cr_in_isvec
.eq(crin_svdec
.isvec
)
1123 comb
+= self
.cr_in_b_isvec
.eq(crin_svdec_b
.isvec
)
1124 comb
+= self
.cr_in_o_isvec
.eq(crin_svdec_o
.isvec
)
1126 # indices are slightly different, BA/BB mess sorted above
1127 comb
+= crin_svdec
.idx
.eq(cr_a_idx
) # SVP64 CR in A
1128 comb
+= crin_svdec_b
.idx
.eq(cr_b_idx
) # SVP64 CR in B
1129 comb
+= crin_svdec_o
.idx
.eq(op
.sv_cr_out
) # SVP64 CR out
1131 # ...and subdecoders' input fields
1132 comb
+= dec_a
.sel_in
.eq(op
.in1_sel
)
1133 comb
+= dec_b
.sel_in
.eq(op
.in2_sel
)
1134 comb
+= dec_c
.sel_in
.eq(op
.in3_sel
)
1135 comb
+= dec_o
.sel_in
.eq(op
.out_sel
)
1136 comb
+= dec_o2
.sel_in
.eq(op
.out_sel
)
1137 if hasattr(do
, "lk"):
1138 comb
+= dec_o2
.lk
.eq(do
.lk
)
1140 # get SVSTATE srcstep (TODO: elwidth, dststep etc.) needed below
1141 srcstep
= Signal
.like(self
.state
.svstate
.srcstep
)
1142 comb
+= srcstep
.eq(self
.state
.svstate
.srcstep
)
1144 # registers a, b, c and out and out2 (LD/ST EA)
1145 for to_reg
, fromreg
, svdec
in (
1146 (e
.read_reg1
, dec_a
.reg_out
, in1_svdec
),
1147 (e
.read_reg2
, dec_b
.reg_out
, in2_svdec
),
1148 (e
.read_reg3
, dec_c
.reg_out
, in3_svdec
),
1149 (e
.write_reg
, dec_o
.reg_out
, o_svdec
),
1150 (e
.write_ea
, dec_o2
.reg_out
, o2_svdec
)):
1151 comb
+= svdec
.extra
.eq(extra
) # EXTRA field of SVP64 RM
1152 comb
+= svdec
.etype
.eq(op
.SV_Etype
) # EXTRA2/3 for this insn
1153 comb
+= svdec
.reg_in
.eq(fromreg
.data
) # 3-bit (CR0/BC/BFA)
1154 comb
+= to_reg
.ok
.eq(fromreg
.ok
)
1155 # detect if Vectorised: add srcstep if yes. TODO: a LOT.
1156 # this trick only holds when elwidth=default and in single-pred
1157 with m
.If(svdec
.isvec
):
1158 comb
+= to_reg
.data
.eq(srcstep
+svdec
.reg_out
) # 7-bit output
1160 comb
+= to_reg
.data
.eq(svdec
.reg_out
) # 7-bit output
1162 comb
+= in1_svdec
.idx
.eq(op
.sv_in1
) # SVP64 reg #1 (matches in1_sel)
1163 comb
+= in2_svdec
.idx
.eq(op
.sv_in2
) # SVP64 reg #2 (matches in2_sel)
1164 comb
+= in3_svdec
.idx
.eq(op
.sv_in3
) # SVP64 reg #3 (matches in3_sel)
1165 comb
+= o_svdec
.idx
.eq(op
.sv_out
) # SVP64 output (matches out_sel)
1166 # XXX TODO - work out where this should come from. the problem is
1167 # that LD-with-update is implied (computed from "is instruction in
1168 # "update mode" rather than specified cleanly as its own CSV column
1169 #comb += o2_svdec.idx.eq(op.sv_out) # SVP64 output (implicit)
1171 # output reg-is-vectorised (and when no output is vectorised)
1172 comb
+= self
.in1_isvec
.eq(in1_svdec
.isvec
)
1173 comb
+= self
.in2_isvec
.eq(in2_svdec
.isvec
)
1174 comb
+= self
.in3_isvec
.eq(in3_svdec
.isvec
)
1175 comb
+= self
.o_isvec
.eq(o_svdec
.isvec
)
1176 comb
+= self
.o2_isvec
.eq(o2_svdec
.isvec
)
1177 # TODO: include SPRs and CRs here! must be True when *all* are scalar
1178 comb
+= self
.no_out_vec
.eq((~o2_svdec
.isvec
) & (~o_svdec
.isvec
))
1181 comb
+= e
.read_spr1
.eq(dec_a
.spr_out
)
1182 comb
+= e
.write_spr
.eq(dec_o
.spr_out
)
1185 comb
+= e
.read_fast1
.eq(dec_a
.fast_out
)
1186 comb
+= e
.read_fast2
.eq(dec_b
.fast_out
)
1187 comb
+= e
.write_fast1
.eq(dec_o
.fast_out
)
1188 comb
+= e
.write_fast2
.eq(dec_o2
.fast_out
)
1190 # condition registers (CR)
1191 for to_reg
, cr
, name
, svdec
in (
1192 (e
.read_cr1
, self
.dec_cr_in
, "cr_bitfield", crin_svdec
),
1193 (e
.read_cr2
, self
.dec_cr_in
, "cr_bitfield_b", crin_svdec_b
),
1194 (e
.read_cr3
, self
.dec_cr_in
, "cr_bitfield_o", crin_svdec_o
),
1195 (e
.write_cr
, self
.dec_cr_out
, "cr_bitfield", crout_svdec
)):
1196 fromreg
= getattr(cr
, name
)
1197 comb
+= svdec
.extra
.eq(extra
) # EXTRA field of SVP64 RM
1198 comb
+= svdec
.etype
.eq(op
.SV_Etype
) # EXTRA2/3 for this insn
1199 comb
+= svdec
.cr_in
.eq(fromreg
.data
) # 3-bit (CR0/BC/BFA)
1200 with m
.If(svdec
.isvec
):
1201 # check if this is CR0 or CR1: treated differently
1202 # (does not "listen" to EXTRA2/3 spec
1203 with m
.If(cr
.sv_override
== 1): # CR0
1204 comb
+= to_reg
.data
.eq(srcstep
+0) # XXX TODO CR0 offset
1205 with m
.Elif(cr
.sv_override
== 2): # CR1
1206 comb
+= to_reg
.data
.eq(srcstep
+1) # XXX TODO CR1 offset
1208 comb
+= to_reg
.data
.eq(srcstep
+svdec
.cr_out
) # 7-bit output
1210 comb
+= to_reg
.data
.eq(svdec
.cr_out
) # 7-bit output
1211 comb
+= to_reg
.ok
.eq(fromreg
.ok
)
1213 # sigh this is exactly the sort of thing for which the
1214 # decoder is designed to not need. MTSPR, MFSPR and others need
1215 # access to the XER bits. however setting e.oe is not appropriate
1216 with m
.If(op
.internal_op
== MicrOp
.OP_MFSPR
):
1217 comb
+= e
.xer_in
.eq(0b111) # SO, CA, OV
1218 with m
.If(op
.internal_op
== MicrOp
.OP_CMP
):
1219 comb
+= e
.xer_in
.eq(1<<XERRegs
.SO
) # SO
1220 with m
.If(op
.internal_op
== MicrOp
.OP_MTSPR
):
1221 comb
+= e
.xer_out
.eq(1)
1223 # set the trapaddr to 0x700 for a td/tw/tdi/twi operation
1224 with m
.If(op
.internal_op
== MicrOp
.OP_TRAP
):
1225 # *DO NOT* call self.trap here. that would reset absolutely
1226 # everything including destroying read of RA and RB.
1227 comb
+= self
.do_copy("trapaddr", 0x70) # strip first nibble
1229 ####################
1230 # ok so the instruction's been decoded, blah blah, however
1231 # now we need to determine if it's actually going to go ahead...
1232 # *or* if in fact it's a privileged operation, whether there's
1233 # an external interrupt, etc. etc. this is a simple priority
1234 # if-elif-elif sequence. decrement takes highest priority,
1235 # EINT next highest, privileged operation third.
1237 # check if instruction is privileged
1238 is_priv_insn
= instr_is_priv(m
, op
.internal_op
, e
.do
.insn
)
1240 # different IRQ conditions
1241 ext_irq_ok
= Signal()
1242 dec_irq_ok
= Signal()
1247 comb
+= ext_irq_ok
.eq(ext_irq
& msr
[MSR
.EE
]) # v3.0B p944 (MSR.EE)
1248 comb
+= dec_irq_ok
.eq(dec_spr
[63] & msr
[MSR
.EE
]) # 6.5.11 p1076
1249 comb
+= priv_ok
.eq(is_priv_insn
& msr
[MSR
.PR
])
1250 comb
+= illeg_ok
.eq(op
.internal_op
== MicrOp
.OP_ILLEGAL
)
1252 # LD/ST exceptions. TestIssuer copies the exception info at us
1253 # after a failed LD/ST.
1254 with m
.If(exc
.happened
):
1255 with m
.If(exc
.alignment
):
1256 self
.trap(m
, TT
.PRIV
, 0x600)
1257 with m
.Elif(exc
.instr_fault
):
1258 with m
.If(exc
.segment_fault
):
1259 self
.trap(m
, TT
.PRIV
, 0x480)
1261 # pass exception info to trap to create SRR1
1262 self
.trap(m
, TT
.MEMEXC
, 0x400, exc
)
1264 with m
.If(exc
.segment_fault
):
1265 self
.trap(m
, TT
.PRIV
, 0x380)
1267 self
.trap(m
, TT
.PRIV
, 0x300)
1269 # decrement counter (v3.0B p1099): TODO 32-bit version (MSR.LPCR)
1270 with m
.Elif(dec_irq_ok
):
1271 self
.trap(m
, TT
.DEC
, 0x900) # v3.0B 6.5 p1065
1273 # external interrupt? only if MSR.EE set
1274 with m
.Elif(ext_irq_ok
):
1275 self
.trap(m
, TT
.EINT
, 0x500)
1277 # privileged instruction trap
1278 with m
.Elif(priv_ok
):
1279 self
.trap(m
, TT
.PRIV
, 0x700)
1281 # illegal instruction must redirect to trap. this is done by
1282 # *overwriting* the decoded instruction and starting again.
1283 # (note: the same goes for interrupts and for privileged operations,
1284 # just with different trapaddr and traptype)
1285 with m
.Elif(illeg_ok
):
1286 # illegal instruction trap
1287 self
.trap(m
, TT
.ILLEG
, 0x700)
1289 # no exception, just copy things to the output
1293 ####################
1294 # follow-up after trap/irq to set up SRR0/1
1296 # trap: (note e.insn_type so this includes OP_ILLEGAL) set up fast regs
1297 # Note: OP_SC could actually be modified to just be a trap
1298 with m
.If((do_out
.insn_type
== MicrOp
.OP_TRAP
) |
1299 (do_out
.insn_type
== MicrOp
.OP_SC
)):
1300 # TRAP write fast1 = SRR0
1301 comb
+= e_out
.write_fast1
.data
.eq(FastRegs
.SRR0
) # constant: SRR0
1302 comb
+= e_out
.write_fast1
.ok
.eq(1)
1303 # TRAP write fast2 = SRR1
1304 comb
+= e_out
.write_fast2
.data
.eq(FastRegs
.SRR1
) # constant: SRR1
1305 comb
+= e_out
.write_fast2
.ok
.eq(1)
1307 # RFID: needs to read SRR0/1
1308 with m
.If(do_out
.insn_type
== MicrOp
.OP_RFID
):
1309 # TRAP read fast1 = SRR0
1310 comb
+= e_out
.read_fast1
.data
.eq(FastRegs
.SRR0
) # constant: SRR0
1311 comb
+= e_out
.read_fast1
.ok
.eq(1)
1312 # TRAP read fast2 = SRR1
1313 comb
+= e_out
.read_fast2
.data
.eq(FastRegs
.SRR1
) # constant: SRR1
1314 comb
+= e_out
.read_fast2
.ok
.eq(1)
1316 # annoying simulator bug
1317 if hasattr(e_out
, "asmcode") and hasattr(self
.dec
.op
, "asmcode"):
1318 comb
+= e_out
.asmcode
.eq(self
.dec
.op
.asmcode
)
1322 def trap(self
, m
, traptype
, trapaddr
, exc
=None):
1323 """trap: this basically "rewrites" the decoded instruction as a trap
1326 op
, e
= self
.dec
.op
, self
.e
1327 comb
+= e
.eq(0) # reset eeeeeverything
1330 comb
+= self
.do_copy("insn", self
.dec
.opcode_in
, True)
1331 comb
+= self
.do_copy("insn_type", MicrOp
.OP_TRAP
, True)
1332 comb
+= self
.do_copy("fn_unit", Function
.TRAP
, True)
1333 comb
+= self
.do_copy("trapaddr", trapaddr
>> 4, True) # bottom 4 bits
1334 comb
+= self
.do_copy("traptype", traptype
, True) # request type
1335 comb
+= self
.do_copy("ldst_exc", exc
, True) # request type
1336 comb
+= self
.do_copy("msr", self
.state
.msr
, True) # copy of MSR "state"
1337 comb
+= self
.do_copy("cia", self
.state
.pc
, True) # copy of PC "state"
1340 # SVP64 Prefix fields: see https://libre-soc.org/openpower/sv/svp64/
1341 # identifies if an instruction is a SVP64-encoded prefix, and extracts
1342 # the 24-bit SVP64 context (RM) if it is
1343 class SVP64PrefixDecoder(Elaboratable
):
1346 self
.opcode_in
= Signal(32, reset_less
=True)
1347 self
.raw_opcode_in
= Signal
.like(self
.opcode_in
, reset_less
=True)
1348 self
.is_svp64_mode
= Signal(1, reset_less
=True)
1349 self
.svp64_rm
= Signal(24, reset_less
=True)
1350 self
.bigendian
= Signal(reset_less
=True)
1352 def elaborate(self
, platform
):
1354 opcode_in
= self
.opcode_in
1356 # sigh copied this from TopPowerDecoder
1357 # raw opcode in assumed to be in LE order: byte-reverse it to get BE
1358 raw_le
= self
.raw_opcode_in
1360 for i
in range(0, 32, 8):
1361 l
.append(raw_le
[i
:i
+8])
1364 comb
+= opcode_in
.eq(Mux(self
.bigendian
, raw_be
, raw_le
))
1366 # start identifying if the incoming opcode is SVP64 prefix)
1367 major
= Signal(6, reset_less
=True)
1368 ident
= Signal(2, reset_less
=True)
1370 comb
+= major
.eq(sel(opcode_in
, SVP64P
.OPC
))
1371 comb
+= ident
.eq(sel(opcode_in
, SVP64P
.SVP64_7_9
))
1373 comb
+= self
.is_svp64_mode
.eq(
1374 (major
== Const(1, 6)) & # EXT01
1375 (ident
== Const(0b11, 2)) # identifier bits
1378 with m
.If(self
.is_svp64_mode
):
1379 # now grab the 24-bit ReMap context bits,
1380 comb
+= self
.svp64_rm
.eq(sel(opcode_in
, SVP64P
.RM
))
1385 return [self
.opcode_in
, self
.raw_opcode_in
, self
.is_svp64_mode
,
1386 self
.svp64_rm
, self
.bigendian
]
1388 def get_rdflags(e
, cu
):
1390 for idx
in range(cu
.n_src
):
1391 regfile
, regname
, _
= cu
.get_in_spec(idx
)
1392 rdflag
, read
= regspec_decode_read(e
, regfile
, regname
)
1394 print("rdflags", rdl
)
1398 if __name__
== '__main__':
1399 svp64
= SVP64PowerDecoder()
1400 vl
= rtlil
.convert(svp64
, ports
=svp64
.ports())
1401 with
open("svp64_dec.il", "w") as f
:
1403 pdecode
= create_pdecode()
1404 dec2
= PowerDecode2(pdecode
)
1405 vl
= rtlil
.convert(dec2
, ports
=dec2
.ports() + pdecode
.ports())
1406 with
open("dec2.il", "w") as f
: