3 not in any way intended for production use. this runs a FSM that:
5 * reads the Program Counter from StateRegs
6 * reads an instruction from a fixed-size Test Memory
7 * issues it to the Simple Core
8 * waits for it to complete
10 * does it all over again
12 the purpose of this module is to verify the functional correctness
13 of the Function Units in the absolute simplest and clearest possible
14 way, and to at provide something that can be further incrementally
18 from nmigen
import (Elaboratable
, Module
, Signal
, ClockSignal
, ResetSignal
,
19 ClockDomain
, DomainRenamer
, Mux
, Const
)
20 from nmigen
.cli
import rtlil
21 from nmigen
.cli
import main
24 from soc
.decoder
.power_decoder
import create_pdecode
25 from soc
.decoder
.power_decoder2
import PowerDecode2
, SVP64PrefixDecoder
26 from soc
.decoder
.decode2execute1
import IssuerDecode2ToOperand
27 from soc
.decoder
.decode2execute1
import Data
28 from soc
.experiment
.testmem
import TestMemory
# test only for instructions
29 from soc
.regfile
.regfiles
import StateRegs
, FastRegs
30 from soc
.simple
.core
import NonProductionCore
31 from soc
.config
.test
.test_loadstore
import TestMemPspec
32 from soc
.config
.ifetch
import ConfigFetchUnit
33 from soc
.decoder
.power_enums
import MicrOp
34 from soc
.debug
.dmi
import CoreDebug
, DMIInterface
35 from soc
.debug
.jtag
import JTAG
36 from soc
.config
.pinouts
import get_pinspecs
37 from soc
.config
.state
import CoreState
38 from soc
.interrupts
.xics
import XICS_ICP
, XICS_ICS
39 from soc
.bus
.simple_gpio
import SimpleGPIO
40 from soc
.bus
.SPBlock512W64B8W
import SPBlock512W64B8W
41 from soc
.clock
.select
import ClockSelect
42 from soc
.clock
.dummypll
import DummyPLL
43 from soc
.sv
.svstate
import SVSTATERec
46 from nmutil
.util
import rising_edge
48 def get_insn(f_instr_o
, pc
):
49 if f_instr_o
.width
== 32:
52 # 64-bit: bit 2 of pc decides which word to select
53 return f_instr_o
.word_select(pc
[2], 32)
55 # gets state input or reads from state regfile
56 def state_get(m
, state_i
, name
, regfile
, regnum
):
60 res
= Signal(64, reset_less
=True, name
=name
)
61 res_ok_delay
= Signal(name
="%s_ok_delay" % name
)
62 sync
+= res_ok_delay
.eq(~state_i
.ok
)
63 with m
.If(state_i
.ok
):
64 # incoming override (start from pc_i)
65 comb
+= res
.eq(state_i
.data
)
67 # otherwise read StateRegs regfile for PC...
68 comb
+= regfile
.ren
.eq(1<<regnum
)
69 # ... but on a 1-clock delay
70 with m
.If(res_ok_delay
):
71 comb
+= res
.eq(regfile
.data_o
)
74 def get_predint(m
, mask
):
75 """decode SVP64 predicate integer mask field to reg number and invert
76 this is identical to the equivalent function in ISACaller except that
77 it doesn't read the INT directly, it just decodes "what needs to be done"
78 i.e. which INT reg, whether it is shifted and whether it is bit-inverted.
84 with m
.Case(SVP64PredInt
.ALWAYS
.value
):
87 with m
.Case(SVP64PredInt
.R3_UNARY
.value
):
90 with m
.Case(SVP64PredInt
.R3
.value
):
92 with m
.Case(SVP64PredInt
.R3_N
.value
):
95 with m
.Case(SVP64PredInt
.R10
.value
):
96 comb
+= regread
.eq(10)
97 with m
.Case(SVP64PredInt
.R10_N
.value
):
98 comb
+= regread
.eq(10)
100 with m
.Case(SVP64PredInt
.R30
.value
):
101 comb
+= regread
.eq(30)
102 with m
.Case(SVP64PredInt
.R30_N
.value
):
103 comb
+= regread
.eq(30)
105 return regread
, invert
, unary
107 def get_predcr(m
, mask
):
108 """decode SVP64 predicate CR to reg number field and invert status
109 this is identical to _get_predcr in ISACaller
114 with m
.Case(SVP64PredCR
.LT
.value
):
117 with m
.Case(SVP64PredCR
.GE
.value
):
120 with m
.Case(SVP64PredCR
.GT
.value
):
123 with m
.Case(SVP64PredCR
.LE
.value
):
126 with m
.Case(SVP64PredCR
.EQ
.value
):
129 with m
.Case(SVP64PredCR
.NE
.value
):
132 with m
.Case(SVP64PredCR
.SO
.value
):
135 with m
.Case(SVP64PredCR
.NS
.value
):
141 class TestIssuerInternal(Elaboratable
):
142 """TestIssuer - reads instructions from TestMemory and issues them
144 efficiency and speed is not the main goal here: functional correctness is.
146 def __init__(self
, pspec
):
148 # test is SVP64 is to be enabled
149 self
.svp64_en
= hasattr(pspec
, "svp64") and (pspec
.svp64
== True)
151 # JTAG interface. add this right at the start because if it's
152 # added it *modifies* the pspec, by adding enable/disable signals
153 # for parts of the rest of the core
154 self
.jtag_en
= hasattr(pspec
, "debug") and pspec
.debug
== 'jtag'
156 subset
= {'uart', 'mtwi', 'eint', 'gpio', 'mspi0', 'mspi1',
158 self
.jtag
= JTAG(get_pinspecs(subset
=subset
))
159 # add signals to pspec to enable/disable icache and dcache
160 # (or data and intstruction wishbone if icache/dcache not included)
161 # https://bugs.libre-soc.org/show_bug.cgi?id=520
162 # TODO: do we actually care if these are not domain-synchronised?
163 # honestly probably not.
164 pspec
.wb_icache_en
= self
.jtag
.wb_icache_en
165 pspec
.wb_dcache_en
= self
.jtag
.wb_dcache_en
166 self
.wb_sram_en
= self
.jtag
.wb_sram_en
168 self
.wb_sram_en
= Const(1)
170 # add 4k sram blocks?
171 self
.sram4x4k
= (hasattr(pspec
, "sram4x4kblock") and
172 pspec
.sram4x4kblock
== True)
176 self
.sram4k
.append(SPBlock512W64B8W(name
="sram4k_%d" % i
,
179 # add interrupt controller?
180 self
.xics
= hasattr(pspec
, "xics") and pspec
.xics
== True
182 self
.xics_icp
= XICS_ICP()
183 self
.xics_ics
= XICS_ICS()
184 self
.int_level_i
= self
.xics_ics
.int_level_i
186 # add GPIO peripheral?
187 self
.gpio
= hasattr(pspec
, "gpio") and pspec
.gpio
== True
189 self
.simple_gpio
= SimpleGPIO()
190 self
.gpio_o
= self
.simple_gpio
.gpio_o
192 # main instruction core. suitable for prototyping / demo only
193 self
.core
= core
= NonProductionCore(pspec
)
195 # instruction decoder. goes into Trap Record
196 pdecode
= create_pdecode()
197 self
.cur_state
= CoreState("cur") # current state (MSR/PC/SVSTATE)
198 self
.pdecode2
= PowerDecode2(pdecode
, state
=self
.cur_state
,
199 opkls
=IssuerDecode2ToOperand
,
200 svp64_en
=self
.svp64_en
)
202 self
.svp64
= SVP64PrefixDecoder() # for decoding SVP64 prefix
204 # Test Instruction memory
205 self
.imem
= ConfigFetchUnit(pspec
).fu
208 self
.dbg
= CoreDebug()
210 # instruction go/monitor
211 self
.pc_o
= Signal(64, reset_less
=True)
212 self
.pc_i
= Data(64, "pc_i") # set "ok" to indicate "please change me"
213 self
.svstate_i
= Data(32, "svstate_i") # ditto
214 self
.core_bigendian_i
= Signal() # TODO: set based on MSR.LE
215 self
.busy_o
= Signal(reset_less
=True)
216 self
.memerr_o
= Signal(reset_less
=True)
218 # STATE regfile read /write ports for PC, MSR, SVSTATE
219 staterf
= self
.core
.regs
.rf
['state']
220 self
.state_r_pc
= staterf
.r_ports
['cia'] # PC rd
221 self
.state_w_pc
= staterf
.w_ports
['d_wr1'] # PC wr
222 self
.state_r_msr
= staterf
.r_ports
['msr'] # MSR rd
223 self
.state_r_sv
= staterf
.r_ports
['sv'] # SVSTATE rd
224 self
.state_w_sv
= staterf
.w_ports
['sv'] # SVSTATE wr
226 # DMI interface access
227 intrf
= self
.core
.regs
.rf
['int']
228 crrf
= self
.core
.regs
.rf
['cr']
229 xerrf
= self
.core
.regs
.rf
['xer']
230 self
.int_r
= intrf
.r_ports
['dmi'] # INT read
231 self
.cr_r
= crrf
.r_ports
['full_cr_dbg'] # CR read
232 self
.xer_r
= xerrf
.r_ports
['full_xer'] # XER read
235 self
.int_pred
= intrf
.r_ports
['pred'] # INT predicate read
236 self
.cr_pred
= crrf
.r_ports
['cr_pred'] # CR predicate read
238 # hack method of keeping an eye on whether branch/trap set the PC
239 self
.state_nia
= self
.core
.regs
.rf
['state'].w_ports
['nia']
240 self
.state_nia
.wen
.name
= 'state_nia_wen'
242 # pulse to synchronize the simulator at instruction end
243 self
.insn_done
= Signal()
246 # store copies of predicate masks
247 self
.srcmask
= Signal(64)
248 self
.dstmask
= Signal(64)
250 def fetch_fsm(self
, m
, core
, pc
, svstate
, nia
, is_svp64_mode
,
251 fetch_pc_ready_o
, fetch_pc_valid_i
,
252 fetch_insn_valid_o
, fetch_insn_ready_i
):
254 this FSM performs fetch of raw instruction data, partial-decodes
255 it 32-bit at a time to detect SVP64 prefixes, and will optionally
256 read a 2nd 32-bit quantity if that occurs.
260 pdecode2
= self
.pdecode2
261 cur_state
= self
.cur_state
262 dec_opcode_i
= pdecode2
.dec
.raw_opcode_in
# raw opcode
264 msr_read
= Signal(reset
=1)
266 with m
.FSM(name
='fetch_fsm'):
269 with m
.State("IDLE"):
270 comb
+= fetch_pc_ready_o
.eq(1)
271 with m
.If(fetch_pc_valid_i
):
272 # instruction allowed to go: start by reading the PC
273 # capture the PC and also drop it into Insn Memory
274 # we have joined a pair of combinatorial memory
275 # lookups together. this is Generally Bad.
276 comb
+= self
.imem
.a_pc_i
.eq(pc
)
277 comb
+= self
.imem
.a_valid_i
.eq(1)
278 comb
+= self
.imem
.f_valid_i
.eq(1)
279 sync
+= cur_state
.pc
.eq(pc
)
280 sync
+= cur_state
.svstate
.eq(svstate
) # and svstate
282 # initiate read of MSR. arrives one clock later
283 comb
+= self
.state_r_msr
.ren
.eq(1 << StateRegs
.MSR
)
284 sync
+= msr_read
.eq(0)
286 m
.next
= "INSN_READ" # move to "wait for bus" phase
288 # dummy pause to find out why simulation is not keeping up
289 with m
.State("INSN_READ"):
290 # one cycle later, msr/sv read arrives. valid only once.
291 with m
.If(~msr_read
):
292 sync
+= msr_read
.eq(1) # yeah don't read it again
293 sync
+= cur_state
.msr
.eq(self
.state_r_msr
.data_o
)
294 with m
.If(self
.imem
.f_busy_o
): # zzz...
295 # busy: stay in wait-read
296 comb
+= self
.imem
.a_valid_i
.eq(1)
297 comb
+= self
.imem
.f_valid_i
.eq(1)
299 # not busy: instruction fetched
300 insn
= get_insn(self
.imem
.f_instr_o
, cur_state
.pc
)
303 # decode the SVP64 prefix, if any
304 comb
+= svp64
.raw_opcode_in
.eq(insn
)
305 comb
+= svp64
.bigendian
.eq(self
.core_bigendian_i
)
306 # pass the decoded prefix (if any) to PowerDecoder2
307 sync
+= pdecode2
.sv_rm
.eq(svp64
.svp64_rm
)
308 # remember whether this is a prefixed instruction, so
309 # the FSM can readily loop when VL==0
310 sync
+= is_svp64_mode
.eq(svp64
.is_svp64_mode
)
311 # calculate the address of the following instruction
312 insn_size
= Mux(svp64
.is_svp64_mode
, 8, 4)
313 sync
+= nia
.eq(cur_state
.pc
+ insn_size
)
314 with m
.If(~svp64
.is_svp64_mode
):
315 # with no prefix, store the instruction
316 # and hand it directly to the next FSM
317 sync
+= dec_opcode_i
.eq(insn
)
318 m
.next
= "INSN_READY"
320 # fetch the rest of the instruction from memory
321 comb
+= self
.imem
.a_pc_i
.eq(cur_state
.pc
+ 4)
322 comb
+= self
.imem
.a_valid_i
.eq(1)
323 comb
+= self
.imem
.f_valid_i
.eq(1)
324 m
.next
= "INSN_READ2"
326 # not SVP64 - 32-bit only
327 sync
+= nia
.eq(cur_state
.pc
+ 4)
328 sync
+= dec_opcode_i
.eq(insn
)
329 m
.next
= "INSN_READY"
331 with m
.State("INSN_READ2"):
332 with m
.If(self
.imem
.f_busy_o
): # zzz...
333 # busy: stay in wait-read
334 comb
+= self
.imem
.a_valid_i
.eq(1)
335 comb
+= self
.imem
.f_valid_i
.eq(1)
337 # not busy: instruction fetched
338 insn
= get_insn(self
.imem
.f_instr_o
, cur_state
.pc
+4)
339 sync
+= dec_opcode_i
.eq(insn
)
340 m
.next
= "INSN_READY"
341 # TODO: probably can start looking at pdecode2.rm_dec
342 # here (or maybe even in INSN_READ state, if svp64_mode
343 # detected, in order to trigger - and wait for - the
346 with m
.State("INSN_READY"):
347 # hand over the instruction, to be decoded
348 comb
+= fetch_insn_valid_o
.eq(1)
349 with m
.If(fetch_insn_ready_i
):
352 def fetch_predicate_fsm(self
, m
, core
, TODO
):
353 """fetch_predicate_fsm - obtains (constructs in the case of CR)
354 src/dest predicate masks
356 https://bugs.libre-soc.org/show_bug.cgi?id=617
357 the predicates can be read here, by using IntRegs r_ports['pred']
358 or CRRegs r_ports['pred']. in the case of CRs it will have to
359 be done through multiple reads, extracting one relevant at a time.
360 later, a faster way would be to use the 32-bit-wide CR port but
361 this is more complex decoding, here. equivalent code used in
362 ISACaller is "from soc.decoder.isa.caller import get_predcr"
366 pdecode2
= self
.pdecode2
367 rm_dec
= pdecode2
.rm_dec
# SVP64RMModeDecode
368 predmode
= rm_dec
.predmode
369 srcpred
, dstpred
= rm_dec
.srcpred
, rm_dec
.dstpred
370 cr_pred
, int_pred
= self
.cr_pred
, self
.int_pred
# read regfiles
371 # if predmode == INT:
372 # INT-src sregread, sinvert, sunary = get_predint(m, srcpred)
373 # INT-dst dregread, dinvert, dunary = get_predint(m, dstpred)
374 # TODO read INT-src and INT-dst into self.srcmask+dstmask
375 # elif predmode == CR:
376 # CR-src sidx, sinvert = get_predcr(m, srcpred)
377 # CR-dst didx, dinvert = get_predcr(m, dstpred)
378 # TODO read CR-src and CR-dst into self.srcmask+dstmask with loop
380 # sync += self.srcmask.eq(-1) # set to all 1s
381 # sync += self.dstmask.eq(-1) # set to all 1s
383 def issue_fsm(self
, m
, core
, pc_changed
, sv_changed
, nia
,
384 dbg
, core_rst
, is_svp64_mode
,
385 fetch_pc_ready_o
, fetch_pc_valid_i
,
386 fetch_insn_valid_o
, fetch_insn_ready_i
,
387 exec_insn_valid_i
, exec_insn_ready_o
,
388 exec_pc_valid_o
, exec_pc_ready_i
):
391 decode / issue FSM. this interacts with the "fetch" FSM
392 through fetch_insn_ready/valid (incoming) and fetch_pc_ready/valid
393 (outgoing). also interacts with the "execute" FSM
394 through exec_insn_ready/valid (outgoing) and exec_pc_ready/valid
396 SVP64 RM prefixes have already been set up by the
397 "fetch" phase, so execute is fairly straightforward.
402 pdecode2
= self
.pdecode2
403 cur_state
= self
.cur_state
406 dec_opcode_i
= pdecode2
.dec
.raw_opcode_in
# raw opcode
408 # for updating svstate (things like srcstep etc.)
409 update_svstate
= Signal() # set this (below) if updating
410 new_svstate
= SVSTATERec("new_svstate")
411 comb
+= new_svstate
.eq(cur_state
.svstate
)
413 with m
.FSM(name
="issue_fsm"):
415 # sync with the "fetch" phase which is reading the instruction
416 # at this point, there is no instruction running, that
417 # could inadvertently update the PC.
418 with m
.State("ISSUE_START"):
419 # wait on "core stop" release, before next fetch
420 # need to do this here, in case we are in a VL==0 loop
421 with m
.If(~dbg
.core_stop_o
& ~core_rst
):
422 comb
+= fetch_pc_valid_i
.eq(1) # tell fetch to start
423 with m
.If(fetch_pc_ready_o
): # fetch acknowledged us
426 # tell core it's stopped, and acknowledge debug handshake
427 comb
+= core
.core_stopped_i
.eq(1)
428 comb
+= dbg
.core_stopped_i
.eq(1)
429 # while stopped, allow updating the PC and SVSTATE
430 with m
.If(self
.pc_i
.ok
):
431 comb
+= self
.state_w_pc
.wen
.eq(1 << StateRegs
.PC
)
432 comb
+= self
.state_w_pc
.data_i
.eq(self
.pc_i
.data
)
433 sync
+= pc_changed
.eq(1)
434 with m
.If(self
.svstate_i
.ok
):
435 comb
+= new_svstate
.eq(self
.svstate_i
.data
)
436 comb
+= update_svstate
.eq(1)
437 sync
+= sv_changed
.eq(1)
439 # decode the instruction when it arrives
440 with m
.State("INSN_WAIT"):
441 comb
+= fetch_insn_ready_i
.eq(1)
442 with m
.If(fetch_insn_valid_o
):
443 # decode the instruction
444 sync
+= core
.e
.eq(pdecode2
.e
)
445 sync
+= core
.state
.eq(cur_state
)
446 sync
+= core
.raw_insn_i
.eq(dec_opcode_i
)
447 sync
+= core
.bigendian_i
.eq(self
.core_bigendian_i
)
448 # set RA_OR_ZERO detection in satellite decoders
449 sync
+= core
.sv_a_nz
.eq(pdecode2
.sv_a_nz
)
450 # loop into ISSUE_START if it's a SVP64 instruction
451 # and VL == 0. this because VL==0 is a for-loop
452 # from 0 to 0 i.e. always, always a NOP.
453 cur_vl
= cur_state
.svstate
.vl
454 with m
.If(is_svp64_mode
& (cur_vl
== 0)):
455 # update the PC before fetching the next instruction
456 # since we are in a VL==0 loop, no instruction was
457 # executed that we could be overwriting
458 comb
+= self
.state_w_pc
.wen
.eq(1 << StateRegs
.PC
)
459 comb
+= self
.state_w_pc
.data_i
.eq(nia
)
460 comb
+= self
.insn_done
.eq(1)
461 m
.next
= "ISSUE_START"
463 m
.next
= "INSN_EXECUTE" # move to "execute"
465 # handshake with execution FSM, move to "wait" once acknowledged
466 with m
.State("INSN_EXECUTE"):
467 # with m.If(is_svp64_mode):
468 # TODO advance src/dst step to "skip" over predicated-out
469 # from self.srcmask and self.dstmask
470 # https://bugs.libre-soc.org/show_bug.cgi?id=617#c3
471 # but still without exceeding VL in either case
472 comb
+= exec_insn_valid_i
.eq(1) # trigger execute
473 with m
.If(exec_insn_ready_o
): # execute acknowledged us
474 m
.next
= "EXECUTE_WAIT"
476 with m
.State("EXECUTE_WAIT"):
477 # wait on "core stop" release, at instruction end
478 # need to do this here, in case we are in a VL>1 loop
479 with m
.If(~dbg
.core_stop_o
& ~core_rst
):
480 comb
+= exec_pc_ready_i
.eq(1)
481 with m
.If(exec_pc_valid_o
):
482 # precalculate srcstep+1 and dststep+1
483 next_srcstep
= Signal
.like(cur_state
.svstate
.srcstep
)
484 next_dststep
= Signal
.like(cur_state
.svstate
.dststep
)
485 comb
+= next_srcstep
.eq(cur_state
.svstate
.srcstep
+1)
486 comb
+= next_dststep
.eq(cur_state
.svstate
.dststep
+1)
488 # was this the last loop iteration?
490 cur_vl
= cur_state
.svstate
.vl
491 comb
+= is_last
.eq(next_srcstep
== cur_vl
)
493 # if either PC or SVSTATE were changed by the previous
494 # instruction, go directly back to Fetch, without
495 # updating either PC or SVSTATE
496 with m
.If(pc_changed | sv_changed
):
497 m
.next
= "ISSUE_START"
499 # also return to Fetch, when no output was a vector
500 # (regardless of SRCSTEP and VL), or when the last
501 # instruction was really the last one of the VL loop
502 with m
.Elif((~pdecode2
.loop_continue
) | is_last
):
503 # before going back to fetch, update the PC state
504 # register with the NIA.
505 # ok here we are not reading the branch unit.
506 # TODO: this just blithely overwrites whatever
507 # pipeline updated the PC
508 comb
+= self
.state_w_pc
.wen
.eq(1 << StateRegs
.PC
)
509 comb
+= self
.state_w_pc
.data_i
.eq(nia
)
510 # reset SRCSTEP before returning to Fetch
511 with m
.If(pdecode2
.loop_continue
):
512 comb
+= new_svstate
.srcstep
.eq(0)
513 comb
+= new_svstate
.dststep
.eq(0)
514 comb
+= update_svstate
.eq(1)
515 m
.next
= "ISSUE_START"
517 # returning to Execute? then, first update SRCSTEP
519 comb
+= new_svstate
.srcstep
.eq(next_srcstep
)
520 comb
+= new_svstate
.dststep
.eq(next_dststep
)
521 comb
+= update_svstate
.eq(1)
525 comb
+= core
.core_stopped_i
.eq(1)
526 comb
+= dbg
.core_stopped_i
.eq(1)
527 # while stopped, allow updating the PC and SVSTATE
528 with m
.If(self
.pc_i
.ok
):
529 comb
+= self
.state_w_pc
.wen
.eq(1 << StateRegs
.PC
)
530 comb
+= self
.state_w_pc
.data_i
.eq(self
.pc_i
.data
)
531 sync
+= pc_changed
.eq(1)
532 with m
.If(self
.svstate_i
.ok
):
533 comb
+= new_svstate
.eq(self
.svstate_i
.data
)
534 comb
+= update_svstate
.eq(1)
535 sync
+= sv_changed
.eq(1)
537 # need to decode the instruction again, after updating SRCSTEP
538 # in the previous state.
539 # mostly a copy of INSN_WAIT, but without the actual wait
540 with m
.State("DECODE_SV"):
541 # decode the instruction
542 sync
+= core
.e
.eq(pdecode2
.e
)
543 sync
+= core
.state
.eq(cur_state
)
544 sync
+= core
.bigendian_i
.eq(self
.core_bigendian_i
)
545 sync
+= core
.sv_a_nz
.eq(pdecode2
.sv_a_nz
)
546 m
.next
= "INSN_EXECUTE" # move to "execute"
548 # check if svstate needs updating: if so, write it to State Regfile
549 with m
.If(update_svstate
):
550 comb
+= self
.state_w_sv
.wen
.eq(1<<StateRegs
.SVSTATE
)
551 comb
+= self
.state_w_sv
.data_i
.eq(new_svstate
)
552 sync
+= cur_state
.svstate
.eq(new_svstate
) # for next clock
554 def execute_fsm(self
, m
, core
, pc_changed
, sv_changed
,
555 exec_insn_valid_i
, exec_insn_ready_o
,
556 exec_pc_valid_o
, exec_pc_ready_i
):
559 execute FSM. this interacts with the "issue" FSM
560 through exec_insn_ready/valid (incoming) and exec_pc_ready/valid
561 (outgoing). SVP64 RM prefixes have already been set up by the
562 "issue" phase, so execute is fairly straightforward.
567 pdecode2
= self
.pdecode2
570 core_busy_o
= core
.busy_o
# core is busy
571 core_ivalid_i
= core
.ivalid_i
# instruction is valid
572 core_issue_i
= core
.issue_i
# instruction is issued
573 insn_type
= core
.e
.do
.insn_type
# instruction MicroOp type
575 with m
.FSM(name
="exec_fsm"):
577 # waiting for instruction bus (stays there until not busy)
578 with m
.State("INSN_START"):
579 comb
+= exec_insn_ready_o
.eq(1)
580 with m
.If(exec_insn_valid_i
):
581 comb
+= core_ivalid_i
.eq(1) # instruction is valid
582 comb
+= core_issue_i
.eq(1) # and issued
583 sync
+= sv_changed
.eq(0)
584 sync
+= pc_changed
.eq(0)
585 m
.next
= "INSN_ACTIVE" # move to "wait completion"
587 # instruction started: must wait till it finishes
588 with m
.State("INSN_ACTIVE"):
589 with m
.If(insn_type
!= MicrOp
.OP_NOP
):
590 comb
+= core_ivalid_i
.eq(1) # instruction is valid
591 # note changes to PC and SVSTATE
592 with m
.If(self
.state_nia
.wen
& (1<<StateRegs
.SVSTATE
)):
593 sync
+= sv_changed
.eq(1)
594 with m
.If(self
.state_nia
.wen
& (1<<StateRegs
.PC
)):
595 sync
+= pc_changed
.eq(1)
596 with m
.If(~core_busy_o
): # instruction done!
597 comb
+= exec_pc_valid_o
.eq(1)
598 with m
.If(exec_pc_ready_i
):
599 comb
+= self
.insn_done
.eq(1)
600 m
.next
= "INSN_START" # back to fetch
602 def setup_peripherals(self
, m
):
603 comb
, sync
= m
.d
.comb
, m
.d
.sync
605 m
.submodules
.core
= core
= DomainRenamer("coresync")(self
.core
)
606 m
.submodules
.imem
= imem
= self
.imem
607 m
.submodules
.dbg
= dbg
= self
.dbg
609 m
.submodules
.jtag
= jtag
= self
.jtag
610 # TODO: UART2GDB mux, here, from external pin
611 # see https://bugs.libre-soc.org/show_bug.cgi?id=499
612 sync
+= dbg
.dmi
.connect_to(jtag
.dmi
)
614 cur_state
= self
.cur_state
616 # 4x 4k SRAM blocks. these simply "exist", they get routed in litex
618 for i
, sram
in enumerate(self
.sram4k
):
619 m
.submodules
["sram4k_%d" % i
] = sram
620 comb
+= sram
.enable
.eq(self
.wb_sram_en
)
622 # XICS interrupt handler
624 m
.submodules
.xics_icp
= icp
= self
.xics_icp
625 m
.submodules
.xics_ics
= ics
= self
.xics_ics
626 comb
+= icp
.ics_i
.eq(ics
.icp_o
) # connect ICS to ICP
627 sync
+= cur_state
.eint
.eq(icp
.core_irq_o
) # connect ICP to core
629 # GPIO test peripheral
631 m
.submodules
.simple_gpio
= simple_gpio
= self
.simple_gpio
633 # connect one GPIO output to ICS bit 15 (like in microwatt soc.vhdl)
634 # XXX causes litex ECP5 test to get wrong idea about input and output
635 # (but works with verilator sim *sigh*)
636 #if self.gpio and self.xics:
637 # comb += self.int_level_i[15].eq(simple_gpio.gpio_o[0])
639 # instruction decoder
640 pdecode
= create_pdecode()
641 m
.submodules
.dec2
= pdecode2
= self
.pdecode2
643 m
.submodules
.svp64
= svp64
= self
.svp64
646 dmi
, d_reg
, d_cr
, d_xer
, = dbg
.dmi
, dbg
.d_gpr
, dbg
.d_cr
, dbg
.d_xer
647 intrf
= self
.core
.regs
.rf
['int']
649 # clock delay power-on reset
650 cd_por
= ClockDomain(reset_less
=True)
651 cd_sync
= ClockDomain()
652 core_sync
= ClockDomain("coresync")
653 m
.domains
+= cd_por
, cd_sync
, core_sync
655 ti_rst
= Signal(reset_less
=True)
656 delay
= Signal(range(4), reset
=3)
657 with m
.If(delay
!= 0):
658 m
.d
.por
+= delay
.eq(delay
- 1)
659 comb
+= cd_por
.clk
.eq(ClockSignal())
661 # power-on reset delay
662 core_rst
= ResetSignal("coresync")
663 comb
+= ti_rst
.eq(delay
!= 0 | dbg
.core_rst_o |
ResetSignal())
664 comb
+= core_rst
.eq(ti_rst
)
666 # busy/halted signals from core
667 comb
+= self
.busy_o
.eq(core
.busy_o
)
668 comb
+= pdecode2
.dec
.bigendian
.eq(self
.core_bigendian_i
)
670 # temporary hack: says "go" immediately for both address gen and ST
672 ldst
= core
.fus
.fus
['ldst0']
673 st_go_edge
= rising_edge(m
, ldst
.st
.rel_o
)
674 m
.d
.comb
+= ldst
.ad
.go_i
.eq(ldst
.ad
.rel_o
) # link addr-go direct to rel
675 m
.d
.comb
+= ldst
.st
.go_i
.eq(st_go_edge
) # link store-go to rising rel
679 def elaborate(self
, platform
):
682 comb
, sync
= m
.d
.comb
, m
.d
.sync
683 cur_state
= self
.cur_state
684 pdecode2
= self
.pdecode2
688 # set up peripherals and core
689 core_rst
= self
.setup_peripherals(m
)
691 # PC and instruction from I-Memory
692 comb
+= self
.pc_o
.eq(cur_state
.pc
)
693 pc_changed
= Signal() # note write to PC
694 sv_changed
= Signal() # note write to SVSTATE
696 # read state either from incoming override or from regfile
697 # TODO: really should be doing MSR in the same way
698 pc
= state_get(m
, self
.pc_i
, "pc", # read PC
699 self
.state_r_pc
, StateRegs
.PC
)
700 svstate
= state_get(m
, self
.svstate_i
, "svstate", # read SVSTATE
701 self
.state_r_sv
, StateRegs
.SVSTATE
)
703 # don't write pc every cycle
704 comb
+= self
.state_w_pc
.wen
.eq(0)
705 comb
+= self
.state_w_pc
.data_i
.eq(0)
707 # don't read msr every cycle
708 comb
+= self
.state_r_msr
.ren
.eq(0)
710 # address of the next instruction, in the absence of a branch
711 # depends on the instruction size
712 nia
= Signal(64, reset_less
=True)
714 # connect up debug signals
715 # TODO comb += core.icache_rst_i.eq(dbg.icache_rst_o)
716 comb
+= dbg
.terminate_i
.eq(core
.core_terminate_o
)
717 comb
+= dbg
.state
.pc
.eq(pc
)
718 comb
+= dbg
.state
.svstate
.eq(svstate
)
719 comb
+= dbg
.state
.msr
.eq(cur_state
.msr
)
721 # pass the prefix mode from Fetch to Issue, so the latter can loop
723 is_svp64_mode
= Signal()
725 # there are *THREE* FSMs, fetch (32/64-bit) issue, decode/execute.
726 # these are the handshake signals between fetch and decode/execute
728 # fetch FSM can run as soon as the PC is valid
729 fetch_pc_valid_i
= Signal() # Execute tells Fetch "start next read"
730 fetch_pc_ready_o
= Signal() # Fetch Tells SVSTATE "proceed"
732 # fetch FSM hands over the instruction to be decoded / issued
733 fetch_insn_valid_o
= Signal()
734 fetch_insn_ready_i
= Signal()
736 # issue FSM delivers the instruction to the be executed
737 exec_insn_valid_i
= Signal()
738 exec_insn_ready_o
= Signal()
740 # execute FSM, hands over the PC/SVSTATE back to the issue FSM
741 exec_pc_valid_o
= Signal()
742 exec_pc_ready_i
= Signal()
744 # the FSMs here are perhaps unusual in that they detect conditions
745 # then "hold" information, combinatorially, for the core
746 # (as opposed to using sync - which would be on a clock's delay)
747 # this includes the actual opcode, valid flags and so on.
749 # Fetch, then Issue, then Execute. Issue is where the VL for-loop
750 # lives. the ready/valid signalling is used to communicate between
753 self
.fetch_fsm(m
, core
, pc
, svstate
, nia
, is_svp64_mode
,
754 fetch_pc_ready_o
, fetch_pc_valid_i
,
755 fetch_insn_valid_o
, fetch_insn_ready_i
)
757 self
.issue_fsm(m
, core
, pc_changed
, sv_changed
, nia
,
758 dbg
, core_rst
, is_svp64_mode
,
759 fetch_pc_ready_o
, fetch_pc_valid_i
,
760 fetch_insn_valid_o
, fetch_insn_ready_i
,
761 exec_insn_valid_i
, exec_insn_ready_o
,
762 exec_pc_valid_o
, exec_pc_ready_i
)
764 self
.execute_fsm(m
, core
, pc_changed
, sv_changed
,
765 exec_insn_valid_i
, exec_insn_ready_o
,
766 exec_pc_valid_o
, exec_pc_ready_i
)
768 # this bit doesn't have to be in the FSM: connect up to read
769 # regfiles on demand from DMI
772 # DEC and TB inc/dec FSM. copy of DEC is put into CoreState,
773 # (which uses that in PowerDecoder2 to raise 0x900 exception)
774 self
.tb_dec_fsm(m
, cur_state
.dec
)
778 def do_dmi(self
, m
, dbg
):
781 dmi
, d_reg
, d_cr
, d_xer
, = dbg
.dmi
, dbg
.d_gpr
, dbg
.d_cr
, dbg
.d_xer
782 intrf
= self
.core
.regs
.rf
['int']
784 with m
.If(d_reg
.req
): # request for regfile access being made
785 # TODO: error-check this
786 # XXX should this be combinatorial? sync better?
788 comb
+= self
.int_r
.ren
.eq(1<<d_reg
.addr
)
790 comb
+= self
.int_r
.addr
.eq(d_reg
.addr
)
791 comb
+= self
.int_r
.ren
.eq(1)
792 d_reg_delay
= Signal()
793 sync
+= d_reg_delay
.eq(d_reg
.req
)
794 with m
.If(d_reg_delay
):
795 # data arrives one clock later
796 comb
+= d_reg
.data
.eq(self
.int_r
.data_o
)
797 comb
+= d_reg
.ack
.eq(1)
799 # sigh same thing for CR debug
800 with m
.If(d_cr
.req
): # request for regfile access being made
801 comb
+= self
.cr_r
.ren
.eq(0b11111111) # enable all
802 d_cr_delay
= Signal()
803 sync
+= d_cr_delay
.eq(d_cr
.req
)
804 with m
.If(d_cr_delay
):
805 # data arrives one clock later
806 comb
+= d_cr
.data
.eq(self
.cr_r
.data_o
)
807 comb
+= d_cr
.ack
.eq(1)
810 with m
.If(d_xer
.req
): # request for regfile access being made
811 comb
+= self
.xer_r
.ren
.eq(0b111111) # enable all
812 d_xer_delay
= Signal()
813 sync
+= d_xer_delay
.eq(d_xer
.req
)
814 with m
.If(d_xer_delay
):
815 # data arrives one clock later
816 comb
+= d_xer
.data
.eq(self
.xer_r
.data_o
)
817 comb
+= d_xer
.ack
.eq(1)
819 def tb_dec_fsm(self
, m
, spr_dec
):
822 this is a FSM for updating either dec or tb. it runs alternately
823 DEC, TB, DEC, TB. note that SPR pipeline could have written a new
824 value to DEC, however the regfile has "passthrough" on it so this
827 see v3.0B p1097-1099 for Timeer Resource and p1065 and p1076
830 comb
, sync
= m
.d
.comb
, m
.d
.sync
831 fast_rf
= self
.core
.regs
.rf
['fast']
832 fast_r_dectb
= fast_rf
.r_ports
['issue'] # DEC/TB
833 fast_w_dectb
= fast_rf
.w_ports
['issue'] # DEC/TB
837 # initiates read of current DEC
838 with m
.State("DEC_READ"):
839 comb
+= fast_r_dectb
.addr
.eq(FastRegs
.DEC
)
840 comb
+= fast_r_dectb
.ren
.eq(1)
843 # waits for DEC read to arrive (1 cycle), updates with new value
844 with m
.State("DEC_WRITE"):
846 # TODO: MSR.LPCR 32-bit decrement mode
847 comb
+= new_dec
.eq(fast_r_dectb
.data_o
- 1)
848 comb
+= fast_w_dectb
.addr
.eq(FastRegs
.DEC
)
849 comb
+= fast_w_dectb
.wen
.eq(1)
850 comb
+= fast_w_dectb
.data_i
.eq(new_dec
)
851 sync
+= spr_dec
.eq(new_dec
) # copy into cur_state for decoder
854 # initiates read of current TB
855 with m
.State("TB_READ"):
856 comb
+= fast_r_dectb
.addr
.eq(FastRegs
.TB
)
857 comb
+= fast_r_dectb
.ren
.eq(1)
860 # waits for read TB to arrive, initiates write of current TB
861 with m
.State("TB_WRITE"):
863 comb
+= new_tb
.eq(fast_r_dectb
.data_o
+ 1)
864 comb
+= fast_w_dectb
.addr
.eq(FastRegs
.TB
)
865 comb
+= fast_w_dectb
.wen
.eq(1)
866 comb
+= fast_w_dectb
.data_i
.eq(new_tb
)
872 yield from self
.pc_i
.ports()
875 yield from self
.core
.ports()
876 yield from self
.imem
.ports()
877 yield self
.core_bigendian_i
883 def external_ports(self
):
884 ports
= self
.pc_i
.ports()
885 ports
+= [self
.pc_o
, self
.memerr_o
, self
.core_bigendian_i
, self
.busy_o
,
889 ports
+= list(self
.jtag
.external_ports())
891 # don't add DMI if JTAG is enabled
892 ports
+= list(self
.dbg
.dmi
.ports())
894 ports
+= list(self
.imem
.ibus
.fields
.values())
895 ports
+= list(self
.core
.l0
.cmpi
.lsmem
.lsi
.slavebus
.fields
.values())
898 for sram
in self
.sram4k
:
899 ports
+= list(sram
.bus
.fields
.values())
902 ports
+= list(self
.xics_icp
.bus
.fields
.values())
903 ports
+= list(self
.xics_ics
.bus
.fields
.values())
904 ports
.append(self
.int_level_i
)
907 ports
+= list(self
.simple_gpio
.bus
.fields
.values())
908 ports
.append(self
.gpio_o
)
916 class TestIssuer(Elaboratable
):
917 def __init__(self
, pspec
):
918 self
.ti
= TestIssuerInternal(pspec
)
920 self
.pll
= DummyPLL()
922 # PLL direct clock or not
923 self
.pll_en
= hasattr(pspec
, "use_pll") and pspec
.use_pll
925 self
.pll_18_o
= Signal(reset_less
=True)
927 def elaborate(self
, platform
):
931 # TestIssuer runs at direct clock
932 m
.submodules
.ti
= ti
= self
.ti
933 cd_int
= ClockDomain("coresync")
936 # ClockSelect runs at PLL output internal clock rate
937 m
.submodules
.pll
= pll
= self
.pll
939 # add clock domains from PLL
940 cd_pll
= ClockDomain("pllclk")
943 # PLL clock established. has the side-effect of running clklsel
944 # at the PLL's speed (see DomainRenamer("pllclk") above)
945 pllclk
= ClockSignal("pllclk")
946 comb
+= pllclk
.eq(pll
.clk_pll_o
)
948 # wire up external 24mhz to PLL
949 comb
+= pll
.clk_24_i
.eq(ClockSignal())
951 # output 18 mhz PLL test signal
952 comb
+= self
.pll_18_o
.eq(pll
.pll_18_o
)
954 # now wire up ResetSignals. don't mind them being in this domain
955 pll_rst
= ResetSignal("pllclk")
956 comb
+= pll_rst
.eq(ResetSignal())
958 # internal clock is set to selector clock-out. has the side-effect of
959 # running TestIssuer at this speed (see DomainRenamer("intclk") above)
960 intclk
= ClockSignal("coresync")
962 comb
+= intclk
.eq(pll
.clk_pll_o
)
964 comb
+= intclk
.eq(ClockSignal())
969 return list(self
.ti
.ports()) + list(self
.pll
.ports()) + \
970 [ClockSignal(), ResetSignal()]
972 def external_ports(self
):
973 ports
= self
.ti
.external_ports()
974 ports
.append(ClockSignal())
975 ports
.append(ResetSignal())
977 ports
.append(self
.pll
.clk_sel_i
)
978 ports
.append(self
.pll_18_o
)
979 ports
.append(self
.pll
.pll_lck_o
)
983 if __name__
== '__main__':
984 units
= {'alu': 1, 'cr': 1, 'branch': 1, 'trap': 1, 'logical': 1,
990 pspec
= TestMemPspec(ldst_ifacetype
='bare_wb',
991 imem_ifacetype
='bare_wb',
996 dut
= TestIssuer(pspec
)
997 vl
= main(dut
, ports
=dut
.ports(), name
="test_issuer")
999 if len(sys
.argv
) == 1:
1000 vl
= rtlil
.convert(dut
, ports
=dut
.external_ports(), name
="test_issuer")
1001 with
open("test_issuer.il", "w") as f
: