3 based on Anton Blanchard microwatt dcache.vhdl
7 from enum
import Enum
, unique
9 from nmigen
import Module
, Signal
, Elaboratable
, Cat
, Repl
, Array
, Const
11 from nmigen
.hdl
.ast
import Display
16 from nmigen
.cli
import main
17 from nmutil
.iocontrol
import RecordObject
18 from nmutil
.util
import wrap
19 from nmigen
.utils
import log2_int
20 from soc
.experiment
.mem_types
import (LoadStore1ToDCacheType
,
21 DCacheToLoadStore1Type
,
25 from soc
.experiment
.wb_types
import (WB_ADDR_BITS
, WB_DATA_BITS
, WB_SEL_BITS
,
26 WBAddrType
, WBDataType
, WBSelType
,
27 WBMasterOut
, WBSlaveOut
,
28 WBMasterOutVector
, WBSlaveOutVector
,
29 WBIOMasterOut
, WBIOSlaveOut
)
31 from soc
.experiment
.cache_ram
import CacheRam
32 from soc
.experiment
.plru
import PLRU
35 from nmigen_soc
.wishbone
.sram
import SRAM
36 from nmigen
import Memory
37 from nmigen
.cli
import rtlil
39 from nmigen
.back
.pysim
import Simulator
, Delay
, Settle
41 from nmigen
.sim
.cxxsim
import Simulator
, Delay
, Settle
44 # TODO: make these parameters of DCache at some point
45 LINE_SIZE
= 64 # Line size in bytes
46 NUM_LINES
= 16 # Number of lines in a set
47 NUM_WAYS
= 4 # Number of ways
48 TLB_SET_SIZE
= 64 # L1 DTLB entries per set
49 TLB_NUM_WAYS
= 2 # L1 DTLB number of sets
50 TLB_LG_PGSZ
= 12 # L1 DTLB log_2(page_size)
51 LOG_LENGTH
= 0 # Non-zero to enable log data collection
53 # BRAM organisation: We never access more than
54 # -- WB_DATA_BITS at a time so to save
55 # -- resources we make the array only that wide, and
56 # -- use consecutive indices for to make a cache "line"
58 # -- ROW_SIZE is the width in bytes of the BRAM
59 # -- (based on WB, so 64-bits)
60 ROW_SIZE
= WB_DATA_BITS
// 8;
62 # ROW_PER_LINE is the number of row (wishbone
63 # transactions) in a line
64 ROW_PER_LINE
= LINE_SIZE
// ROW_SIZE
66 # BRAM_ROWS is the number of rows in BRAM needed
67 # to represent the full dcache
68 BRAM_ROWS
= NUM_LINES
* ROW_PER_LINE
71 # Bit fields counts in the address
73 # REAL_ADDR_BITS is the number of real address
77 # ROW_BITS is the number of bits to select a row
78 ROW_BITS
= log2_int(BRAM_ROWS
)
80 # ROW_LINE_BITS is the number of bits to select
82 ROW_LINE_BITS
= log2_int(ROW_PER_LINE
)
84 # LINE_OFF_BITS is the number of bits for
85 # the offset in a cache line
86 LINE_OFF_BITS
= log2_int(LINE_SIZE
)
88 # ROW_OFF_BITS is the number of bits for
90 ROW_OFF_BITS
= log2_int(ROW_SIZE
)
92 # INDEX_BITS is the number if bits to
94 INDEX_BITS
= log2_int(NUM_LINES
)
96 # SET_SIZE_BITS is the log base 2 of the set size
97 SET_SIZE_BITS
= LINE_OFF_BITS
+ INDEX_BITS
99 # TAG_BITS is the number of bits of
100 # the tag part of the address
101 TAG_BITS
= REAL_ADDR_BITS
- SET_SIZE_BITS
103 # TAG_WIDTH is the width in bits of each way of the tag RAM
104 TAG_WIDTH
= TAG_BITS
+ 7 - ((TAG_BITS
+ 7) % 8)
106 # WAY_BITS is the number of bits to select a way
107 WAY_BITS
= log2_int(NUM_WAYS
)
109 # Example of layout for 32 lines of 64 bytes:
111 # .. tag |index| line |
113 # .. | |---| | ROW_LINE_BITS (3)
114 # .. | |--- - --| LINE_OFF_BITS (6)
115 # .. | |- --| ROW_OFF_BITS (3)
116 # .. |----- ---| | ROW_BITS (8)
117 # .. |-----| | INDEX_BITS (5)
118 # .. --------| | TAG_BITS (45)
120 TAG_RAM_WIDTH
= TAG_WIDTH
* NUM_WAYS
123 return Array(Signal(TAG_RAM_WIDTH
) for x
in range(NUM_LINES
))
125 def CacheValidBitsArray():
126 return Array(Signal(INDEX_BITS
) for x
in range(NUM_LINES
))
128 def RowPerLineValidArray():
129 return Array(Signal(name
="rows_valid%d" % x
) for x
in range(ROW_PER_LINE
))
132 TLB_SET_BITS
= log2_int(TLB_SET_SIZE
)
133 TLB_WAY_BITS
= log2_int(TLB_NUM_WAYS
)
134 TLB_EA_TAG_BITS
= 64 - (TLB_LG_PGSZ
+ TLB_SET_BITS
)
135 TLB_TAG_WAY_BITS
= TLB_NUM_WAYS
* TLB_EA_TAG_BITS
137 TLB_PTE_WAY_BITS
= TLB_NUM_WAYS
* TLB_PTE_BITS
;
139 assert (LINE_SIZE
% ROW_SIZE
) == 0, "LINE_SIZE not multiple of ROW_SIZE"
140 assert (LINE_SIZE
% 2) == 0, "LINE_SIZE not power of 2"
141 assert (NUM_LINES
% 2) == 0, "NUM_LINES not power of 2"
142 assert (ROW_PER_LINE
% 2) == 0, "ROW_PER_LINE not power of 2"
143 assert ROW_BITS
== (INDEX_BITS
+ ROW_LINE_BITS
), "geometry bits don't add up"
144 assert (LINE_OFF_BITS
== ROW_OFF_BITS
+ ROW_LINE_BITS
), \
145 "geometry bits don't add up"
146 assert REAL_ADDR_BITS
== (TAG_BITS
+ INDEX_BITS
+ LINE_OFF_BITS
), \
147 "geometry bits don't add up"
148 assert REAL_ADDR_BITS
== (TAG_BITS
+ ROW_BITS
+ ROW_OFF_BITS
), \
149 "geometry bits don't add up"
150 assert 64 == WB_DATA_BITS
, "Can't yet handle wb width that isn't 64-bits"
151 assert SET_SIZE_BITS
<= TLB_LG_PGSZ
, "Set indexed by virtual address"
154 def TLBValidBitsArray():
155 return Array(Signal(TLB_NUM_WAYS
) for x
in range(TLB_SET_SIZE
))
158 return Array(Signal(TLB_EA_TAG_BITS
) for x
in range (TLB_NUM_WAYS
))
161 return Array(Signal(TLB_TAG_WAY_BITS
) for x
in range (TLB_SET_SIZE
))
164 return Array(Signal(TLB_PTE_WAY_BITS
) for x
in range(TLB_SET_SIZE
))
167 return Array(Signal(WAY_BITS
) for x
in range(TLB_NUM_WAYS
))
169 # Cache RAM interface
171 return Array(Signal(WB_DATA_BITS
) for x
in range(NUM_WAYS
))
173 # PLRU output interface
175 return Array(Signal(WAY_BITS
) for x
in range(NUM_LINES
))
177 # TLB PLRU output interface
179 return Array(Signal(TLB_WAY_BITS
) for x
in range(TLB_SET_SIZE
))
181 # Helper functions to decode incoming requests
183 # Return the cache line index (tag index) for an address
185 return addr
[LINE_OFF_BITS
:SET_SIZE_BITS
]
187 # Return the cache row index (data memory) for an address
189 return addr
[ROW_OFF_BITS
:SET_SIZE_BITS
]
191 # Return the index of a row within a line
192 def get_row_of_line(row
):
193 return row
[:ROW_LINE_BITS
]
195 # Returns whether this is the last row of a line
196 def is_last_row_addr(addr
, last
):
197 return addr
[ROW_OFF_BITS
:LINE_OFF_BITS
] == last
199 # Returns whether this is the last row of a line
200 def is_last_row(row
, last
):
201 return get_row_of_line(row
) == last
203 # Return the next row in the current cache line. We use a
204 # dedicated function in order to limit the size of the
205 # generated adder to be only the bits within a cache line
206 # (3 bits with default settings)
208 row_v
= row
[0:ROW_LINE_BITS
] + 1
209 return Cat(row_v
[:ROW_LINE_BITS
], row
[ROW_LINE_BITS
:])
211 # Get the tag value from the address
213 return addr
[SET_SIZE_BITS
:REAL_ADDR_BITS
]
215 # Read a tag from a tag memory row
216 def read_tag(way
, tagset
):
217 return tagset
.word_select(way
, TAG_WIDTH
)[:TAG_BITS
]
219 # Read a TLB tag from a TLB tag memory row
220 def read_tlb_tag(way
, tags
):
221 return tags
.word_select(way
, TLB_EA_TAG_BITS
)
223 # Write a TLB tag to a TLB tag memory row
224 def write_tlb_tag(way
, tags
, tag
):
225 return read_tlb_tag(way
, tags
).eq(tag
)
227 # Read a PTE from a TLB PTE memory row
228 def read_tlb_pte(way
, ptes
):
229 return ptes
.word_select(way
, TLB_PTE_BITS
)
231 def write_tlb_pte(way
, ptes
, newpte
):
232 return read_tlb_pte(way
, ptes
).eq(newpte
)
235 # Record for storing permission, attribute, etc. bits from a PTE
236 class PermAttr(RecordObject
):
237 def __init__(self
, name
=None):
238 super().__init
__(name
=name
)
239 self
.reference
= Signal()
240 self
.changed
= Signal()
241 self
.nocache
= Signal()
243 self
.rd_perm
= Signal()
244 self
.wr_perm
= Signal()
247 def extract_perm_attr(pte
):
249 pa
.reference
= pte
[8]
258 # Type of operation on a "valid" input
262 OP_BAD
= 1 # NC cache hit, TLB miss, prot/RC failure
263 OP_STCX_FAIL
= 2 # conditional store w/o reservation
264 OP_LOAD_HIT
= 3 # Cache hit on load
265 OP_LOAD_MISS
= 4 # Load missing cache
266 OP_LOAD_NC
= 5 # Non-cachable load
267 OP_STORE_HIT
= 6 # Store hitting cache
268 OP_STORE_MISS
= 7 # Store missing cache
271 # Cache state machine
274 IDLE
= 0 # Normal load hit processing
275 RELOAD_WAIT_ACK
= 1 # Cache reload wait ack
276 STORE_WAIT_ACK
= 2 # Store wait ack
277 NC_LOAD_WAIT_ACK
= 3 # Non-cachable load wait ack
282 # In order to make timing, we use the BRAMs with
283 # an output buffer, which means that the BRAM
284 # output is delayed by an extra cycle.
286 # Thus, the dcache has a 2-stage internal pipeline
287 # for cache hits with no stalls.
289 # All other operations are handled via stalling
290 # in the first stage.
292 # The second stage can thus complete a hit at the same
293 # time as the first stage emits a stall for a complex op.
295 # Stage 0 register, basically contains just the latched request
297 class RegStage0(RecordObject
):
298 def __init__(self
, name
=None):
299 super().__init
__(name
=name
)
300 self
.req
= LoadStore1ToDCacheType(name
="lsmem")
301 self
.tlbie
= Signal()
302 self
.doall
= Signal()
303 self
.tlbld
= Signal()
304 self
.mmu_req
= Signal() # indicates source of request
307 class MemAccessRequest(RecordObject
):
308 def __init__(self
, name
=None):
309 super().__init
__(name
=name
)
311 self
.valid
= Signal()
313 self
.real_addr
= Signal(REAL_ADDR_BITS
)
314 self
.data
= Signal(64)
315 self
.byte_sel
= Signal(8)
316 self
.hit_way
= Signal(WAY_BITS
)
317 self
.same_tag
= Signal()
318 self
.mmu_req
= Signal()
321 # First stage register, contains state for stage 1 of load hits
322 # and for the state machine used by all other operations
323 class RegStage1(RecordObject
):
324 def __init__(self
, name
=None):
325 super().__init
__(name
=name
)
326 # Info about the request
327 self
.full
= Signal() # have uncompleted request
328 self
.mmu_req
= Signal() # request is from MMU
329 self
.req
= MemAccessRequest(name
="reqmem")
332 self
.hit_way
= Signal(WAY_BITS
)
333 self
.hit_load_valid
= Signal()
334 self
.hit_index
= Signal(INDEX_BITS
)
335 self
.cache_hit
= Signal()
338 self
.tlb_hit
= Signal()
339 self
.tlb_hit_way
= Signal(TLB_NUM_WAYS
)
340 self
.tlb_hit_index
= Signal(TLB_WAY_BITS
)
342 # 2-stage data buffer for data forwarded from writes to reads
343 self
.forward_data1
= Signal(64)
344 self
.forward_data2
= Signal(64)
345 self
.forward_sel1
= Signal(8)
346 self
.forward_valid1
= Signal()
347 self
.forward_way1
= Signal(WAY_BITS
)
348 self
.forward_row1
= Signal(ROW_BITS
)
349 self
.use_forward1
= Signal()
350 self
.forward_sel
= Signal(8)
352 # Cache miss state (reload state machine)
353 self
.state
= Signal(State
)
355 self
.write_bram
= Signal()
356 self
.write_tag
= Signal()
357 self
.slow_valid
= Signal()
358 self
.wb
= WBMasterOut()
359 self
.reload_tag
= Signal(TAG_BITS
)
360 self
.store_way
= Signal(WAY_BITS
)
361 self
.store_row
= Signal(ROW_BITS
)
362 self
.store_index
= Signal(INDEX_BITS
)
363 self
.end_row_ix
= Signal(ROW_LINE_BITS
)
364 self
.rows_valid
= RowPerLineValidArray()
365 self
.acks_pending
= Signal(3)
366 self
.inc_acks
= Signal()
367 self
.dec_acks
= Signal()
369 # Signals to complete (possibly with error)
370 self
.ls_valid
= Signal()
371 self
.ls_error
= Signal()
372 self
.mmu_done
= Signal()
373 self
.mmu_error
= Signal()
374 self
.cache_paradox
= Signal()
376 # Signal to complete a failed stcx.
377 self
.stcx_fail
= Signal()
380 # Reservation information
381 class Reservation(RecordObject
):
384 self
.valid
= Signal()
385 self
.addr
= Signal(64-LINE_OFF_BITS
)
388 class DTLBUpdate(Elaboratable
):
390 self
.tlbie
= Signal()
391 self
.tlbwe
= Signal()
392 self
.doall
= Signal()
393 self
.updated
= Signal()
394 self
.v_updated
= Signal()
395 self
.tlb_hit
= Signal()
396 self
.tlb_req_index
= Signal(TLB_SET_BITS
)
398 self
.tlb_hit_way
= Signal(TLB_WAY_BITS
)
399 self
.tlb_tag_way
= Signal(TLB_TAG_WAY_BITS
)
400 self
.tlb_pte_way
= Signal(TLB_PTE_WAY_BITS
)
401 self
.repl_way
= Signal(TLB_WAY_BITS
)
402 self
.eatag
= Signal(TLB_EA_TAG_BITS
)
403 self
.pte_data
= Signal(TLB_PTE_BITS
)
405 self
.dv
= Signal(TLB_PTE_WAY_BITS
)
407 self
.tb_out
= Signal(TLB_TAG_WAY_BITS
)
408 self
.pb_out
= Signal(TLB_NUM_WAYS
)
409 self
.db_out
= Signal(TLB_PTE_WAY_BITS
)
411 def elaborate(self
, platform
):
416 tagset
= Signal(TLB_TAG_WAY_BITS
)
417 pteset
= Signal(TLB_PTE_WAY_BITS
)
419 tb_out
, pb_out
, db_out
= self
.tb_out
, self
.pb_out
, self
.db_out
421 with m
.If(self
.tlbie
& self
.doall
):
422 pass # clear all back in parent
423 with m
.Elif(self
.tlbie
):
424 with m
.If(self
.tlb_hit
):
425 comb
+= db_out
.eq(self
.dv
)
426 comb
+= db_out
.bit_select(self
.tlb_hit_way
, 1).eq(1)
427 comb
+= self
.v_updated
.eq(1)
429 with m
.Elif(self
.tlbwe
):
431 comb
+= tagset
.eq(self
.tlb_tag_way
)
432 comb
+= write_tlb_tag(self
.repl_way
, tagset
, self
.eatag
)
433 comb
+= tb_out
.eq(tagset
)
435 comb
+= pteset
.eq(self
.tlb_pte_way
)
436 comb
+= write_tlb_pte(self
.repl_way
, pteset
, self
.pte_data
)
437 comb
+= pb_out
.eq(pteset
)
439 comb
+= db_out
.bit_select(self
.repl_way
, 1).eq(1)
441 comb
+= self
.updated
.eq(1)
442 comb
+= self
.v_updated
.eq(1)
446 def dcache_request(self
, m
, r0
, ra
, req_index
, req_row
, req_tag
,
447 r0_valid
, r1
, cache_valid_bits
, replace_way
,
448 use_forward1_next
, use_forward2_next
,
449 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
450 valid_ra
, perm_ok
, access_ok
, req_op
, req_go
,
452 tlb_hit
, tlb_hit_way
, tlb_valid_way
, cache_tag_set
,
453 cancel_store
, req_same_tag
, r0_stall
, early_req_row
):
454 """Cache request parsing and hit detection
457 class DCachePendingHit(Elaboratable
):
459 def __init__(self
, tlb_pte_way
, tlb_valid_way
, tlb_hit_way
,
460 cache_valid_idx
, cache_tag_set
,
465 self
.virt_mode
= Signal()
466 self
.is_hit
= Signal()
467 self
.tlb_hit
= Signal()
468 self
.hit_way
= Signal(WAY_BITS
)
469 self
.rel_match
= Signal()
470 self
.req_index
= Signal(INDEX_BITS
)
471 self
.reload_tag
= Signal(TAG_BITS
)
473 self
.tlb_hit_way
= tlb_hit_way
474 self
.tlb_pte_way
= tlb_pte_way
475 self
.tlb_valid_way
= tlb_valid_way
476 self
.cache_valid_idx
= cache_valid_idx
477 self
.cache_tag_set
= cache_tag_set
478 self
.req_addr
= req_addr
479 self
.hit_set
= hit_set
481 def elaborate(self
, platform
):
487 virt_mode
= self
.virt_mode
489 tlb_pte_way
= self
.tlb_pte_way
490 tlb_valid_way
= self
.tlb_valid_way
491 cache_valid_idx
= self
.cache_valid_idx
492 cache_tag_set
= self
.cache_tag_set
493 req_addr
= self
.req_addr
494 tlb_hit_way
= self
.tlb_hit_way
495 tlb_hit
= self
.tlb_hit
496 hit_set
= self
.hit_set
497 hit_way
= self
.hit_way
498 rel_match
= self
.rel_match
499 req_index
= self
.req_index
500 reload_tag
= self
.reload_tag
502 rel_matches
= Array(Signal() for i
in range(TLB_NUM_WAYS
))
503 hit_way_set
= HitWaySet()
505 # Test if pending request is a hit on any way
506 # In order to make timing in virtual mode,
507 # when we are using the TLB, we compare each
508 # way with each of the real addresses from each way of
509 # the TLB, and then decide later which match to use.
511 with m
.If(virt_mode
):
512 for j
in range(TLB_NUM_WAYS
):
513 s_tag
= Signal(TAG_BITS
, name
="s_tag%d" % j
)
515 s_pte
= Signal(TLB_PTE_BITS
)
516 s_ra
= Signal(REAL_ADDR_BITS
)
517 comb
+= s_pte
.eq(read_tlb_pte(j
, tlb_pte_way
))
518 comb
+= s_ra
.eq(Cat(req_addr
[0:TLB_LG_PGSZ
],
519 s_pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
520 comb
+= s_tag
.eq(get_tag(s_ra
))
522 for i
in range(NUM_WAYS
):
523 is_tag_hit
= Signal(name
="is_tag_hit_%d_%d" % (j
, i
))
524 comb
+= is_tag_hit
.eq(go
& cache_valid_idx
[i
] &
525 (read_tag(i
, cache_tag_set
) == s_tag
)
527 with m
.If(is_tag_hit
):
528 comb
+= hit_way_set
[j
].eq(i
)
530 comb
+= hit_set
[j
].eq(s_hit
)
531 with m
.If(s_tag
== reload_tag
):
532 comb
+= rel_matches
[j
].eq(1)
534 comb
+= is_hit
.eq(hit_set
[tlb_hit_way
])
535 comb
+= hit_way
.eq(hit_way_set
[tlb_hit_way
])
536 comb
+= rel_match
.eq(rel_matches
[tlb_hit_way
])
538 s_tag
= Signal(TAG_BITS
)
539 comb
+= s_tag
.eq(get_tag(req_addr
))
540 for i
in range(NUM_WAYS
):
541 is_tag_hit
= Signal(name
="is_tag_hit_%d" % i
)
542 comb
+= is_tag_hit
.eq(go
& cache_valid_idx
[i
] &
543 (read_tag(i
, cache_tag_set
) == s_tag
))
544 with m
.If(is_tag_hit
):
545 comb
+= hit_way
.eq(i
)
547 with m
.If(s_tag
== reload_tag
):
548 comb
+= rel_match
.eq(1)
553 class DCache(Elaboratable
):
554 """Set associative dcache write-through
555 TODO (in no specific order):
556 * See list in icache.vhdl
557 * Complete load misses on the cycle when WB data comes instead of
558 at the end of line (this requires dealing with requests coming in
562 self
.d_in
= LoadStore1ToDCacheType("d_in")
563 self
.d_out
= DCacheToLoadStore1Type("d_out")
565 self
.m_in
= MMUToDCacheType("m_in")
566 self
.m_out
= DCacheToMMUType("m_out")
568 self
.stall_out
= Signal()
570 self
.wb_out
= WBMasterOut()
571 self
.wb_in
= WBSlaveOut()
573 self
.log_out
= Signal(20)
575 def stage_0(self
, m
, r0
, r1
, r0_full
):
576 """Latch the request in r0.req as long as we're not stalling
580 d_in
, d_out
, m_in
= self
.d_in
, self
.d_out
, self
.m_in
582 r
= RegStage0("stage0")
584 # TODO, this goes in unit tests and formal proofs
585 with m
.If(~
(d_in
.valid
& m_in
.valid
)):
586 #sync += Display("request collision loadstore vs MMU")
589 with m
.If(m_in
.valid
):
590 sync
+= r
.req
.valid
.eq(1)
591 sync
+= r
.req
.load
.eq(~
(m_in
.tlbie | m_in
.tlbld
))
592 sync
+= r
.req
.dcbz
.eq(0)
593 sync
+= r
.req
.nc
.eq(0)
594 sync
+= r
.req
.reserve
.eq(0)
595 sync
+= r
.req
.virt_mode
.eq(1)
596 sync
+= r
.req
.priv_mode
.eq(1)
597 sync
+= r
.req
.addr
.eq(m_in
.addr
)
598 sync
+= r
.req
.data
.eq(m_in
.pte
)
599 sync
+= r
.req
.byte_sel
.eq(~
0) # Const -1 sets all to 0b111....
600 sync
+= r
.tlbie
.eq(m_in
.tlbie
)
601 sync
+= r
.doall
.eq(m_in
.doall
)
602 sync
+= r
.tlbld
.eq(m_in
.tlbld
)
603 sync
+= r
.mmu_req
.eq(1)
605 sync
+= r
.req
.eq(d_in
)
606 sync
+= r
.tlbie
.eq(0)
607 sync
+= r
.doall
.eq(0)
608 sync
+= r
.tlbld
.eq(0)
609 sync
+= r
.mmu_req
.eq(0)
610 with m
.If(~
(r1
.full
& r0_full
)):
612 sync
+= r0_full
.eq(r
.req
.valid
)
614 def tlb_read(self
, m
, r0_stall
, tlb_valid_way
,
615 tlb_tag_way
, tlb_pte_way
, dtlb_valid_bits
,
616 dtlb_tags
, dtlb_ptes
):
618 Operates in the second cycle on the request latched in r0.req.
619 TLB updates write the entry at the end of the second cycle.
623 m_in
, d_in
= self
.m_in
, self
.d_in
625 index
= Signal(TLB_SET_BITS
)
626 addrbits
= Signal(TLB_SET_BITS
)
629 amax
= TLB_LG_PGSZ
+ TLB_SET_BITS
631 with m
.If(m_in
.valid
):
632 comb
+= addrbits
.eq(m_in
.addr
[amin
: amax
])
634 comb
+= addrbits
.eq(d_in
.addr
[amin
: amax
])
635 comb
+= index
.eq(addrbits
)
637 # If we have any op and the previous op isn't finished,
638 # then keep the same output for next cycle.
639 with m
.If(~r0_stall
):
640 sync
+= tlb_valid_way
.eq(dtlb_valid_bits
[index
])
641 sync
+= tlb_tag_way
.eq(dtlb_tags
[index
])
642 sync
+= tlb_pte_way
.eq(dtlb_ptes
[index
])
644 def maybe_tlb_plrus(self
, m
, r1
, tlb_plru_victim
):
645 """Generate TLB PLRUs
650 if TLB_NUM_WAYS
== 0:
652 for i
in range(TLB_SET_SIZE
):
654 tlb_plru
= PLRU(WAY_BITS
)
655 setattr(m
.submodules
, "maybe_plru_%d" % i
, tlb_plru
)
656 tlb_plru_acc_en
= Signal()
658 comb
+= tlb_plru_acc_en
.eq(r1
.tlb_hit
& (r1
.tlb_hit_index
== i
))
659 comb
+= tlb_plru
.acc_en
.eq(tlb_plru_acc_en
)
660 comb
+= tlb_plru
.acc
.eq(r1
.tlb_hit_way
)
661 comb
+= tlb_plru_victim
[i
].eq(tlb_plru
.lru_o
)
663 def tlb_search(self
, m
, tlb_req_index
, r0
, r0_valid
,
664 tlb_valid_way
, tlb_tag_way
, tlb_hit_way
,
665 tlb_pte_way
, pte
, tlb_hit
, valid_ra
, perm_attr
, ra
):
670 hitway
= Signal(TLB_WAY_BITS
)
672 eatag
= Signal(TLB_EA_TAG_BITS
)
674 TLB_LG_END
= TLB_LG_PGSZ
+ TLB_SET_BITS
675 comb
+= tlb_req_index
.eq(r0
.req
.addr
[TLB_LG_PGSZ
: TLB_LG_END
])
676 comb
+= eatag
.eq(r0
.req
.addr
[TLB_LG_END
: 64 ])
678 for i
in range(TLB_NUM_WAYS
):
679 is_tag_hit
= Signal()
680 comb
+= is_tag_hit
.eq(tlb_valid_way
[i
]
681 & read_tlb_tag(i
, tlb_tag_way
) == eatag
)
682 with m
.If(is_tag_hit
):
686 comb
+= tlb_hit
.eq(hit
& r0_valid
)
687 comb
+= tlb_hit_way
.eq(hitway
)
690 comb
+= pte
.eq(read_tlb_pte(hitway
, tlb_pte_way
))
693 comb
+= valid_ra
.eq(tlb_hit | ~r0
.req
.virt_mode
)
694 with m
.If(r0
.req
.virt_mode
):
695 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
696 r0
.req
.addr
[ROW_OFF_BITS
:TLB_LG_PGSZ
],
697 pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
698 comb
+= perm_attr
.eq(extract_perm_attr(pte
))
700 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
701 r0
.req
.addr
[ROW_OFF_BITS
:REAL_ADDR_BITS
]))
703 comb
+= perm_attr
.reference
.eq(1)
704 comb
+= perm_attr
.changed
.eq(1)
705 comb
+= perm_attr
.nocache
.eq(0)
706 comb
+= perm_attr
.priv
.eq(1)
707 comb
+= perm_attr
.rd_perm
.eq(1)
708 comb
+= perm_attr
.wr_perm
.eq(1)
710 def tlb_update(self
, m
, r0_valid
, r0
, dtlb_valid_bits
, tlb_req_index
,
711 tlb_hit_way
, tlb_hit
, tlb_plru_victim
, tlb_tag_way
,
712 dtlb_tags
, tlb_pte_way
, dtlb_ptes
):
720 comb
+= tlbie
.eq(r0_valid
& r0
.tlbie
)
721 comb
+= tlbwe
.eq(r0_valid
& r0
.tlbld
)
723 m
.submodules
.tlb_update
= d
= DTLBUpdate()
724 with m
.If(tlbie
& r0
.doall
):
725 # clear all valid bits at once
726 for i
in range(TLB_SET_SIZE
):
727 sync
+= dtlb_valid_bits
[i
].eq(0)
728 with m
.If(d
.updated
):
729 sync
+= dtlb_tags
[tlb_req_index
].eq(d
.tb_out
)
730 sync
+= dtlb_ptes
[tlb_req_index
].eq(d
.pb_out
)
731 with m
.If(d
.v_updated
):
732 sync
+= dtlb_valid_bits
[tlb_req_index
].eq(d
.db_out
)
734 comb
+= d
.dv
.eq(dtlb_valid_bits
[tlb_req_index
])
736 comb
+= d
.tlbie
.eq(tlbie
)
737 comb
+= d
.tlbwe
.eq(tlbwe
)
738 comb
+= d
.doall
.eq(r0
.doall
)
739 comb
+= d
.tlb_hit
.eq(tlb_hit
)
740 comb
+= d
.tlb_hit_way
.eq(tlb_hit_way
)
741 comb
+= d
.tlb_tag_way
.eq(tlb_tag_way
)
742 comb
+= d
.tlb_pte_way
.eq(tlb_pte_way
)
743 comb
+= d
.tlb_req_index
.eq(tlb_req_index
)
746 comb
+= d
.repl_way
.eq(tlb_hit_way
)
748 comb
+= d
.repl_way
.eq(tlb_plru_victim
[tlb_req_index
])
749 comb
+= d
.eatag
.eq(r0
.req
.addr
[TLB_LG_PGSZ
+ TLB_SET_BITS
:64])
750 comb
+= d
.pte_data
.eq(r0
.req
.data
)
752 def maybe_plrus(self
, m
, r1
, plru_victim
):
758 if TLB_NUM_WAYS
== 0:
761 for i
in range(NUM_LINES
):
763 plru
= PLRU(WAY_BITS
)
764 setattr(m
.submodules
, "plru%d" % i
, plru
)
765 plru_acc_en
= Signal()
767 comb
+= plru_acc_en
.eq(r1
.cache_hit
& (r1
.hit_index
== i
))
768 comb
+= plru
.acc_en
.eq(plru_acc_en
)
769 comb
+= plru
.acc
.eq(r1
.hit_way
)
770 comb
+= plru_victim
[i
].eq(plru
.lru_o
)
772 def cache_tag_read(self
, m
, r0_stall
, req_index
, cache_tag_set
, cache_tags
):
773 """Cache tag RAM read port
777 m_in
, d_in
= self
.m_in
, self
.d_in
779 index
= Signal(INDEX_BITS
)
782 comb
+= index
.eq(req_index
)
783 with m
.Elif(m_in
.valid
):
784 comb
+= index
.eq(get_index(m_in
.addr
))
786 comb
+= index
.eq(get_index(d_in
.addr
))
787 sync
+= cache_tag_set
.eq(cache_tags
[index
])
789 def dcache_request(self
, m
, r0
, ra
, req_index
, req_row
, req_tag
,
790 r0_valid
, r1
, cache_valid_bits
, replace_way
,
791 use_forward1_next
, use_forward2_next
,
792 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
793 valid_ra
, perm_ok
, access_ok
, req_op
, req_go
,
795 tlb_hit
, tlb_hit_way
, tlb_valid_way
, cache_tag_set
,
796 cancel_store
, req_same_tag
, r0_stall
, early_req_row
):
797 """Cache request parsing and hit detection
802 m_in
, d_in
= self
.m_in
, self
.d_in
805 hit_way
= Signal(WAY_BITS
)
810 hit_set
= Array(Signal() for i
in range(TLB_NUM_WAYS
))
811 cache_valid_idx
= Signal(INDEX_BITS
)
813 # Extract line, row and tag from request
814 comb
+= req_index
.eq(get_index(r0
.req
.addr
))
815 comb
+= req_row
.eq(get_row(r0
.req
.addr
))
816 comb
+= req_tag
.eq(get_tag(ra
))
818 comb
+= go
.eq(r0_valid
& ~
(r0
.tlbie | r0
.tlbld
) & ~r1
.ls_error
)
819 comb
+= cache_valid_idx
.eq(cache_valid_bits
[req_index
])
821 m
.submodules
.dcache_pend
= dc
= DCachePendingHit(tlb_pte_way
,
822 tlb_valid_way
, tlb_hit_way
,
823 cache_valid_idx
, cache_tag_set
,
827 comb
+= dc
.tlb_hit
.eq(tlb_hit
)
828 comb
+= dc
.reload_tag
.eq(r1
.reload_tag
)
829 comb
+= dc
.virt_mode
.eq(r0
.req
.virt_mode
)
831 comb
+= dc
.req_index
.eq(req_index
)
832 comb
+= is_hit
.eq(dc
.is_hit
)
833 comb
+= hit_way
.eq(dc
.hit_way
)
834 comb
+= req_same_tag
.eq(dc
.rel_match
)
836 # See if the request matches the line currently being reloaded
837 with m
.If((r1
.state
== State
.RELOAD_WAIT_ACK
) &
838 (req_index
== r1
.store_index
) & req_same_tag
):
839 # For a store, consider this a hit even if the row isn't
840 # valid since it will be by the time we perform the store.
841 # For a load, check the appropriate row valid bit.
842 valid
= r1
.rows_valid
[req_row
% ROW_PER_LINE
]
843 comb
+= is_hit
.eq(~r0
.req
.load | valid
)
844 comb
+= hit_way
.eq(replace_way
)
846 # Whether to use forwarded data for a load or not
847 comb
+= use_forward1_next
.eq(0)
848 with m
.If((get_row(r1
.req
.real_addr
) == req_row
) &
849 (r1
.req
.hit_way
== hit_way
)):
850 # Only need to consider r1.write_bram here, since if we
851 # are writing refill data here, then we don't have a
852 # cache hit this cycle on the line being refilled.
853 # (There is the possibility that the load following the
854 # load miss that started the refill could be to the old
855 # contents of the victim line, since it is a couple of
856 # cycles after the refill starts before we see the updated
857 # cache tag. In that case we don't use the bypass.)
858 comb
+= use_forward1_next
.eq(r1
.write_bram
)
859 comb
+= use_forward2_next
.eq(0)
860 with m
.If((r1
.forward_row1
== req_row
) & (r1
.forward_way1
== hit_way
)):
861 comb
+= use_forward2_next
.eq(r1
.forward_valid1
)
863 # The way that matched on a hit
864 comb
+= req_hit_way
.eq(hit_way
)
866 # The way to replace on a miss
867 with m
.If(r1
.write_tag
):
868 comb
+= replace_way
.eq(plru_victim
[r1
.store_index
])
870 comb
+= replace_way
.eq(r1
.store_way
)
872 # work out whether we have permission for this access
873 # NB we don't yet implement AMR, thus no KUAP
874 comb
+= rc_ok
.eq(perm_attr
.reference
875 & (r0
.req
.load | perm_attr
.changed
)
877 comb
+= perm_ok
.eq((r0
.req
.priv_mode | ~perm_attr
.priv
) &
879 (r0
.req
.load
& perm_attr
.rd_perm
)))
880 comb
+= access_ok
.eq(valid_ra
& perm_ok
& rc_ok
)
881 # Combine the request and cache hit status to decide what
882 # operation needs to be done
883 comb
+= nc
.eq(r0
.req
.nc | perm_attr
.nocache
)
884 comb
+= op
.eq(Op
.OP_NONE
)
886 with m
.If(~access_ok
):
887 comb
+= op
.eq(Op
.OP_BAD
)
888 with m
.Elif(cancel_store
):
889 comb
+= op
.eq(Op
.OP_STCX_FAIL
)
891 comb
+= opsel
.eq(Cat(is_hit
, nc
, r0
.req
.load
))
892 with m
.Switch(opsel
):
894 comb
+= op
.eq(Op
.OP_LOAD_HIT
)
896 comb
+= op
.eq(Op
.OP_LOAD_MISS
)
898 comb
+= op
.eq(Op
.OP_LOAD_NC
)
900 comb
+= op
.eq(Op
.OP_STORE_HIT
)
902 comb
+= op
.eq(Op
.OP_STORE_MISS
)
904 comb
+= op
.eq(Op
.OP_STORE_MISS
)
906 comb
+= op
.eq(Op
.OP_BAD
)
908 comb
+= op
.eq(Op
.OP_BAD
)
910 comb
+= op
.eq(Op
.OP_NONE
)
911 comb
+= req_op
.eq(op
)
912 comb
+= req_go
.eq(go
)
914 # Version of the row number that is valid one cycle earlier
915 # in the cases where we need to read the cache data BRAM.
916 # If we're stalling then we need to keep reading the last
918 with m
.If(~r0_stall
):
919 with m
.If(m_in
.valid
):
920 comb
+= early_req_row
.eq(get_row(m_in
.addr
))
922 comb
+= early_req_row
.eq(get_row(d_in
.addr
))
924 comb
+= early_req_row
.eq(req_row
)
926 def reservation_comb(self
, m
, cancel_store
, set_rsrv
, clear_rsrv
,
927 r0_valid
, r0
, reservation
):
928 """Handle load-with-reservation and store-conditional instructions
933 with m
.If(r0_valid
& r0
.req
.reserve
):
935 # XXX generate alignment interrupt if address
936 # is not aligned XXX or if r0.req.nc = '1'
937 with m
.If(r0
.req
.load
):
938 comb
+= set_rsrv
.eq(1) # load with reservation
940 comb
+= clear_rsrv
.eq(1) # store conditional
941 with m
.If(~reservation
.valid | r0
.req
.addr
[LINE_OFF_BITS
:64]):
942 comb
+= cancel_store
.eq(1)
944 def reservation_reg(self
, m
, r0_valid
, access_ok
, set_rsrv
, clear_rsrv
,
950 with m
.If(r0_valid
& access_ok
):
951 with m
.If(clear_rsrv
):
952 sync
+= reservation
.valid
.eq(0)
953 with m
.Elif(set_rsrv
):
954 sync
+= reservation
.valid
.eq(1)
955 sync
+= reservation
.addr
.eq(r0
.req
.addr
[LINE_OFF_BITS
:64])
957 def writeback_control(self
, m
, r1
, cache_out
):
958 """Return data for loads & completion control logic
962 d_out
, m_out
= self
.d_out
, self
.m_out
964 data_out
= Signal(64)
965 data_fwd
= Signal(64)
967 # Use the bypass if are reading the row that was
968 # written 1 or 2 cycles ago, including for the
969 # slow_valid = 1 case (i.e. completing a load
970 # miss or a non-cacheable load).
971 with m
.If(r1
.use_forward1
):
972 comb
+= data_fwd
.eq(r1
.forward_data1
)
974 comb
+= data_fwd
.eq(r1
.forward_data2
)
976 comb
+= data_out
.eq(cache_out
[r1
.hit_way
])
979 with m
.If(r1
.forward_sel
[i
]):
980 dsel
= data_fwd
.word_select(i
, 8)
981 comb
+= data_out
.word_select(i
, 8).eq(dsel
)
983 comb
+= d_out
.valid
.eq(r1
.ls_valid
)
984 comb
+= d_out
.data
.eq(data_out
)
985 comb
+= d_out
.store_done
.eq(~r1
.stcx_fail
)
986 comb
+= d_out
.error
.eq(r1
.ls_error
)
987 comb
+= d_out
.cache_paradox
.eq(r1
.cache_paradox
)
990 comb
+= m_out
.done
.eq(r1
.mmu_done
)
991 comb
+= m_out
.err
.eq(r1
.mmu_error
)
992 comb
+= m_out
.data
.eq(data_out
)
994 # We have a valid load or store hit or we just completed
995 # a slow op such as a load miss, a NC load or a store
997 # Note: the load hit is delayed by one cycle. However it
998 # can still not collide with r.slow_valid (well unless I
999 # miscalculated) because slow_valid can only be set on a
1000 # subsequent request and not on its first cycle (the state
1001 # machine must have advanced), which makes slow_valid
1002 # at least 2 cycles from the previous hit_load_valid.
1004 # Sanity: Only one of these must be set in any given cycle
1006 if False: # TODO: need Display to get this to work
1007 assert (r1
.slow_valid
& r1
.stcx_fail
) != 1, \
1008 "unexpected slow_valid collision with stcx_fail"
1010 assert ((r1
.slow_valid | r1
.stcx_fail
) | r1
.hit_load_valid
) != 1, \
1011 "unexpected hit_load_delayed collision with slow_valid"
1013 with m
.If(~r1
.mmu_req
):
1014 # Request came from loadstore1...
1015 # Load hit case is the standard path
1016 with m
.If(r1
.hit_load_valid
):
1017 #Display(f"completing load hit data={data_out}")
1020 # error cases complete without stalling
1021 with m
.If(r1
.ls_error
):
1022 # Display("completing ld/st with error")
1025 # Slow ops (load miss, NC, stores)
1026 with m
.If(r1
.slow_valid
):
1027 #Display(f"completing store or load miss data={data_out}")
1031 # Request came from MMU
1032 with m
.If(r1
.hit_load_valid
):
1033 # Display(f"completing load hit to MMU, data={m_out.data}")
1035 # error cases complete without stalling
1036 with m
.If(r1
.mmu_error
):
1037 #Display("combpleting MMU ld with error")
1040 # Slow ops (i.e. load miss)
1041 with m
.If(r1
.slow_valid
):
1042 #Display("completing MMU load miss, data={m_out.data}")
1045 def rams(self
, m
, r1
, early_req_row
, cache_out
, replace_way
):
1047 Generate a cache RAM for each way. This handles the normal
1048 reads, writes from reloads and the special store-hit update
1051 Note: the BRAMs have an extra read buffer, meaning the output
1052 is pipelined an extra cycle. This differs from the
1053 icache. The writeback logic needs to take that into
1054 account by using 1-cycle delayed signals for load hits.
1059 for i
in range(NUM_WAYS
):
1061 rd_addr
= Signal(ROW_BITS
)
1063 wr_addr
= Signal(ROW_BITS
)
1064 wr_data
= Signal(WB_DATA_BITS
)
1065 wr_sel
= Signal(ROW_SIZE
)
1066 wr_sel_m
= Signal(ROW_SIZE
)
1067 _d_out
= Signal(WB_DATA_BITS
)
1069 way
= CacheRam(ROW_BITS
, WB_DATA_BITS
, True)
1070 setattr(m
.submodules
, "cacheram_%d" % i
, way
)
1072 comb
+= way
.rd_en
.eq(do_read
)
1073 comb
+= way
.rd_addr
.eq(rd_addr
)
1074 comb
+= _d_out
.eq(way
.rd_data_o
)
1075 comb
+= way
.wr_sel
.eq(wr_sel_m
)
1076 comb
+= way
.wr_addr
.eq(wr_addr
)
1077 comb
+= way
.wr_data
.eq(wr_data
)
1080 comb
+= do_read
.eq(1)
1081 comb
+= rd_addr
.eq(early_req_row
)
1082 comb
+= cache_out
[i
].eq(_d_out
)
1086 # Defaults to wishbone read responses (cache refill)
1088 # For timing, the mux on wr_data/sel/addr is not
1089 # dependent on anything other than the current state.
1091 with m
.If(r1
.write_bram
):
1092 # Write store data to BRAM. This happens one
1093 # cycle after the store is in r0.
1094 comb
+= wr_data
.eq(r1
.req
.data
)
1095 comb
+= wr_sel
.eq(r1
.req
.byte_sel
)
1096 comb
+= wr_addr
.eq(get_row(r1
.req
.real_addr
))
1098 with m
.If(i
== r1
.req
.hit_way
):
1099 comb
+= do_write
.eq(1)
1101 # Otherwise, we might be doing a reload or a DCBZ
1103 comb
+= wr_data
.eq(0)
1105 comb
+= wr_data
.eq(wb_in
.dat
)
1106 comb
+= wr_addr
.eq(r1
.store_row
)
1107 comb
+= wr_sel
.eq(~
0) # all 1s
1109 with m
.If((r1
.state
== State
.RELOAD_WAIT_ACK
)
1110 & wb_in
.ack
& (replace_way
== i
)):
1111 comb
+= do_write
.eq(1)
1113 # Mask write selects with do_write since BRAM
1114 # doesn't have a global write-enable
1115 with m
.If(do_write
):
1116 comb
+= wr_sel_m
.eq(wr_sel
)
1118 # Cache hit synchronous machine for the easy case.
1119 # This handles load hits.
1120 # It also handles error cases (TLB miss, cache paradox)
1121 def dcache_fast_hit(self
, m
, req_op
, r0_valid
, r0
, r1
,
1122 req_hit_way
, req_index
, req_tag
, access_ok
,
1123 tlb_hit
, tlb_hit_way
, tlb_req_index
):
1128 with m
.If(req_op
!= Op
.OP_NONE
):
1129 sync
+= Display("op:%d addr:%x nc: %d idx: %x tag: %x way: %x",
1130 req_op
, r0
.req
.addr
, r0
.req
.nc
,
1131 req_index
, req_tag
, req_hit_way
)
1133 with m
.If(r0_valid
):
1134 sync
+= r1
.mmu_req
.eq(r0
.mmu_req
)
1136 # Fast path for load/store hits.
1137 # Set signals for the writeback controls.
1138 sync
+= r1
.hit_way
.eq(req_hit_way
)
1139 sync
+= r1
.hit_index
.eq(req_index
)
1141 with m
.If(req_op
== Op
.OP_LOAD_HIT
):
1142 sync
+= r1
.hit_load_valid
.eq(1)
1144 sync
+= r1
.hit_load_valid
.eq(0)
1146 with m
.If((req_op
== Op
.OP_LOAD_HIT
) |
(req_op
== Op
.OP_STORE_HIT
)):
1147 sync
+= r1
.cache_hit
.eq(1)
1149 sync
+= r1
.cache_hit
.eq(0)
1151 with m
.If(req_op
== Op
.OP_BAD
):
1152 # Display(f"Signalling ld/st error valid_ra={valid_ra}"
1153 # f"rc_ok={rc_ok} perm_ok={perm_ok}"
1154 sync
+= r1
.ls_error
.eq(~r0
.mmu_req
)
1155 sync
+= r1
.mmu_error
.eq(r0
.mmu_req
)
1156 sync
+= r1
.cache_paradox
.eq(access_ok
)
1159 sync
+= r1
.ls_error
.eq(0)
1160 sync
+= r1
.mmu_error
.eq(0)
1161 sync
+= r1
.cache_paradox
.eq(0)
1163 with m
.If(req_op
== Op
.OP_STCX_FAIL
):
1166 sync
+= r1
.stcx_fail
.eq(0)
1168 # Record TLB hit information for updating TLB PLRU
1169 sync
+= r1
.tlb_hit
.eq(tlb_hit
)
1170 sync
+= r1
.tlb_hit_way
.eq(tlb_hit_way
)
1171 sync
+= r1
.tlb_hit_index
.eq(tlb_req_index
)
1173 # Memory accesses are handled by this state machine:
1175 # * Cache load miss/reload (in conjunction with "rams")
1176 # * Load hits for non-cachable forms
1177 # * Stores (the collision case is handled in "rams")
1179 # All wishbone requests generation is done here.
1180 # This machine operates at stage 1.
1181 def dcache_slow(self
, m
, r1
, use_forward1_next
, use_forward2_next
,
1182 cache_valid_bits
, r0
, replace_way
,
1183 req_hit_way
, req_same_tag
,
1184 r0_valid
, req_op
, cache_tag
, req_go
, ra
):
1190 req
= MemAccessRequest("mreq_ds")
1192 adjust_acks
= Signal(3)
1193 stbs_done
= Signal()
1195 sync
+= r1
.use_forward1
.eq(use_forward1_next
)
1196 sync
+= r1
.forward_sel
.eq(0)
1198 with m
.If(use_forward1_next
):
1199 sync
+= r1
.forward_sel
.eq(r1
.req
.byte_sel
)
1200 with m
.Elif(use_forward2_next
):
1201 sync
+= r1
.forward_sel
.eq(r1
.forward_sel1
)
1203 sync
+= r1
.forward_data2
.eq(r1
.forward_data1
)
1204 with m
.If(r1
.write_bram
):
1205 sync
+= r1
.forward_data1
.eq(r1
.req
.data
)
1206 sync
+= r1
.forward_sel1
.eq(r1
.req
.byte_sel
)
1207 sync
+= r1
.forward_way1
.eq(r1
.req
.hit_way
)
1208 sync
+= r1
.forward_row1
.eq(get_row(r1
.req
.real_addr
))
1209 sync
+= r1
.forward_valid1
.eq(1)
1212 sync
+= r1
.forward_data1
.eq(0)
1214 sync
+= r1
.forward_data1
.eq(wb_in
.dat
)
1215 sync
+= r1
.forward_sel1
.eq(~
0) # all 1s
1216 sync
+= r1
.forward_way1
.eq(replace_way
)
1217 sync
+= r1
.forward_row1
.eq(r1
.store_row
)
1218 sync
+= r1
.forward_valid1
.eq(0)
1220 # One cycle pulses reset
1221 sync
+= r1
.slow_valid
.eq(0)
1222 sync
+= r1
.write_bram
.eq(0)
1223 sync
+= r1
.inc_acks
.eq(0)
1224 sync
+= r1
.dec_acks
.eq(0)
1226 sync
+= r1
.ls_valid
.eq(0)
1227 # complete tlbies and TLB loads in the third cycle
1228 sync
+= r1
.mmu_done
.eq(r0_valid
& (r0
.tlbie | r0
.tlbld
))
1230 with m
.If((req_op
== Op
.OP_LOAD_HIT
)
1231 |
(req_op
== Op
.OP_STCX_FAIL
)):
1232 with m
.If(~r0
.mmu_req
):
1233 sync
+= r1
.ls_valid
.eq(1)
1235 sync
+= r1
.mmu_done
.eq(1)
1237 with m
.If(r1
.write_tag
):
1238 # Store new tag in selected way
1239 for i
in range(NUM_WAYS
):
1240 with m
.If(i
== replace_way
):
1241 ct
= Signal(TAG_RAM_WIDTH
)
1242 comb
+= ct
.eq(cache_tag
[r1
.store_index
])
1243 comb
+= ct
.word_select(i
, TAG_WIDTH
).eq(r1
.reload_tag
)
1244 sync
+= cache_tag
[r1
.store_index
].eq(ct
)
1245 sync
+= r1
.store_way
.eq(replace_way
)
1246 sync
+= r1
.write_tag
.eq(0)
1248 # Take request from r1.req if there is one there,
1249 # else from req_op, ra, etc.
1251 comb
+= req
.eq(r1
.req
)
1253 comb
+= req
.op
.eq(req_op
)
1254 comb
+= req
.valid
.eq(req_go
)
1255 comb
+= req
.mmu_req
.eq(r0
.mmu_req
)
1256 comb
+= req
.dcbz
.eq(r0
.req
.dcbz
)
1257 comb
+= req
.real_addr
.eq(ra
)
1259 with m
.If(~r0
.req
.dcbz
):
1260 comb
+= req
.data
.eq(r0
.req
.data
)
1262 comb
+= req
.data
.eq(0)
1264 # Select all bytes for dcbz
1265 # and for cacheable loads
1266 with m
.If(r0
.req
.dcbz |
(r0
.req
.load
& ~r0
.req
.nc
)):
1267 comb
+= req
.byte_sel
.eq(~
0) # all 1s
1269 comb
+= req
.byte_sel
.eq(r0
.req
.byte_sel
)
1270 comb
+= req
.hit_way
.eq(req_hit_way
)
1271 comb
+= req
.same_tag
.eq(req_same_tag
)
1273 # Store the incoming request from r0,
1274 # if it is a slow request
1275 # Note that r1.full = 1 implies req_op = OP_NONE
1276 with m
.If((req_op
== Op
.OP_LOAD_MISS
)
1277 |
(req_op
== Op
.OP_LOAD_NC
)
1278 |
(req_op
== Op
.OP_STORE_MISS
)
1279 |
(req_op
== Op
.OP_STORE_HIT
)):
1280 sync
+= r1
.req
.eq(req
)
1281 sync
+= r1
.full
.eq(1)
1283 # Main state machine
1284 with m
.Switch(r1
.state
):
1286 with m
.Case(State
.IDLE
):
1287 # XXX check 'left downto. probably means len(r1.wb.adr)
1288 # r1.wb.adr <= req.real_addr(
1289 # r1.wb.adr'left downto 0
1291 sync
+= r1
.wb
.adr
.eq(req
.real_addr
)
1292 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1293 sync
+= r1
.wb
.dat
.eq(req
.data
)
1294 sync
+= r1
.dcbz
.eq(req
.dcbz
)
1296 # Keep track of our index and way
1297 # for subsequent stores.
1298 sync
+= r1
.store_index
.eq(get_index(req
.real_addr
))
1299 sync
+= r1
.store_row
.eq(get_row(req
.real_addr
))
1300 sync
+= r1
.end_row_ix
.eq(
1301 get_row_of_line(get_row(req
.real_addr
))
1303 sync
+= r1
.reload_tag
.eq(get_tag(req
.real_addr
))
1304 sync
+= r1
.req
.same_tag
.eq(1)
1306 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1307 sync
+= r1
.store_way
.eq(req
.hit_way
)
1309 # Reset per-row valid bits,
1310 # ready for handling OP_LOAD_MISS
1311 for i
in range(ROW_PER_LINE
):
1312 sync
+= r1
.rows_valid
[i
].eq(0)
1314 with m
.If(req_op
!= Op
.OP_NONE
):
1315 sync
+= Display("cache op %d", req
.op
)
1317 with m
.Switch(req
.op
):
1318 with m
.Case(Op
.OP_LOAD_HIT
):
1319 # stay in IDLE state
1322 with m
.Case(Op
.OP_LOAD_MISS
):
1323 #Display(f"cache miss real addr:" \
1324 # f"{req_real_addr}" \
1325 # f" idx:{get_index(req_real_addr)}" \
1326 # f" tag:{get_tag(req.real_addr)}")
1329 # Start the wishbone cycle
1330 sync
+= r1
.wb
.we
.eq(0)
1331 sync
+= r1
.wb
.cyc
.eq(1)
1332 sync
+= r1
.wb
.stb
.eq(1)
1334 # Track that we had one request sent
1335 sync
+= r1
.state
.eq(State
.RELOAD_WAIT_ACK
)
1336 sync
+= r1
.write_tag
.eq(1)
1338 with m
.Case(Op
.OP_LOAD_NC
):
1339 sync
+= r1
.wb
.cyc
.eq(1)
1340 sync
+= r1
.wb
.stb
.eq(1)
1341 sync
+= r1
.wb
.we
.eq(0)
1342 sync
+= r1
.state
.eq(State
.NC_LOAD_WAIT_ACK
)
1344 with m
.Case(Op
.OP_STORE_HIT
, Op
.OP_STORE_MISS
):
1345 with m
.If(~req
.dcbz
):
1346 sync
+= r1
.state
.eq(State
.STORE_WAIT_ACK
)
1347 sync
+= r1
.acks_pending
.eq(1)
1348 sync
+= r1
.full
.eq(0)
1349 sync
+= r1
.slow_valid
.eq(1)
1351 with m
.If(~req
.mmu_req
):
1352 sync
+= r1
.ls_valid
.eq(1)
1354 sync
+= r1
.mmu_done
.eq(1)
1356 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1357 sync
+= r1
.write_bram
.eq(1)
1359 sync
+= r1
.state
.eq(State
.RELOAD_WAIT_ACK
)
1361 with m
.If(req
.op
== Op
.OP_STORE_MISS
):
1362 sync
+= r1
.write_tag
.eq(1)
1364 sync
+= r1
.wb
.we
.eq(1)
1365 sync
+= r1
.wb
.cyc
.eq(1)
1366 sync
+= r1
.wb
.stb
.eq(1)
1368 # OP_NONE and OP_BAD do nothing
1369 # OP_BAD & OP_STCX_FAIL were
1370 # handled above already
1371 with m
.Case(Op
.OP_NONE
):
1373 with m
.Case(Op
.OP_BAD
):
1375 with m
.Case(Op
.OP_STCX_FAIL
):
1378 with m
.Case(State
.RELOAD_WAIT_ACK
):
1379 # Requests are all sent if stb is 0
1380 comb
+= stbs_done
.eq(~r1
.wb
.stb
)
1382 with m
.If(~wb_in
.stall
& ~stbs_done
):
1383 # That was the last word?
1384 # We are done sending.
1385 # Clear stb and set stbs_done
1386 # so we can handle an eventual
1387 # last ack on the same cycle.
1388 with m
.If(is_last_row_addr(
1389 r1
.wb
.adr
, r1
.end_row_ix
)):
1390 sync
+= r1
.wb
.stb
.eq(0)
1391 comb
+= stbs_done
.eq(0)
1393 # Calculate the next row address in the current cache line
1394 rarange
= r1
.wb
.adr
[ROW_OFF_BITS
: LINE_OFF_BITS
]
1395 sync
+= rarange
.eq(rarange
+ 1)
1397 # Incoming acks processing
1398 sync
+= r1
.forward_valid1
.eq(wb_in
.ack
)
1399 with m
.If(wb_in
.ack
):
1400 # XXX needs an Array bit-accessor here
1401 sync
+= r1
.rows_valid
[r1
.store_row
% ROW_PER_LINE
].eq(1)
1403 # If this is the data we were looking for,
1404 # we can complete the request next cycle.
1405 # Compare the whole address in case the
1406 # request in r1.req is not the one that
1407 # started this refill.
1408 with m
.If(r1
.full
& r1
.req
.same_tag
&
1409 ((r1
.dcbz
& r1
.req
.dcbz
) |
1410 (~r1
.dcbz
& (r1
.req
.op
== Op
.OP_LOAD_MISS
))) &
1411 (r1
.store_row
== get_row(r1
.req
.real_addr
))):
1412 sync
+= r1
.full
.eq(0)
1413 sync
+= r1
.slow_valid
.eq(1)
1414 with m
.If(~r1
.mmu_req
):
1415 sync
+= r1
.ls_valid
.eq(1)
1417 sync
+= r1
.mmu_done
.eq(1)
1418 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1419 sync
+= r1
.use_forward1
.eq(1)
1421 # Check for completion
1422 with m
.If(stbs_done
& is_last_row(r1
.store_row
,
1424 # Complete wishbone cycle
1425 sync
+= r1
.wb
.cyc
.eq(0)
1427 # Cache line is now valid
1428 cv
= Signal(INDEX_BITS
)
1429 sync
+= cv
.eq(cache_valid_bits
[r1
.store_index
])
1430 sync
+= cv
.bit_select(r1
.store_way
, 1).eq(1)
1431 sync
+= r1
.state
.eq(State
.IDLE
)
1433 # Increment store row counter
1434 sync
+= r1
.store_row
.eq(next_row(r1
.store_row
))
1436 with m
.Case(State
.STORE_WAIT_ACK
):
1437 comb
+= stbs_done
.eq(~r1
.wb
.stb
)
1438 comb
+= acks
.eq(r1
.acks_pending
)
1440 with m
.If(r1
.inc_acks
!= r1
.dec_acks
):
1441 with m
.If(r1
.inc_acks
):
1442 comb
+= adjust_acks
.eq(acks
+ 1)
1444 comb
+= adjust_acks
.eq(acks
- 1)
1446 comb
+= adjust_acks
.eq(acks
)
1448 sync
+= r1
.acks_pending
.eq(adjust_acks
)
1450 # Clear stb when slave accepted request
1451 with m
.If(~wb_in
.stall
):
1452 # See if there is another store waiting
1453 # to be done which is in the same real page.
1454 with m
.If(req
.valid
):
1455 ra
= req
.real_addr
[0:SET_SIZE_BITS
]
1456 sync
+= r1
.wb
.adr
[0:SET_SIZE_BITS
].eq(ra
)
1457 sync
+= r1
.wb
.dat
.eq(req
.data
)
1458 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1460 with m
.Elif((adjust_acks
< 7) & req
.same_tag
&
1461 ((req
.op
== Op
.OP_STORE_MISS
)
1462 |
(req
.op
== Op
.OP_STORE_HIT
))):
1463 sync
+= r1
.wb
.stb
.eq(1)
1464 comb
+= stbs_done
.eq(0)
1466 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1467 sync
+= r1
.write_bram
.eq(1)
1468 sync
+= r1
.full
.eq(0)
1469 sync
+= r1
.slow_valid
.eq(1)
1471 # Store requests never come from the MMU
1472 sync
+= r1
.ls_valid
.eq(1)
1473 comb
+= stbs_done
.eq(0)
1474 sync
+= r1
.inc_acks
.eq(1)
1476 sync
+= r1
.wb
.stb
.eq(0)
1477 comb
+= stbs_done
.eq(1)
1479 # Got ack ? See if complete.
1480 with m
.If(wb_in
.ack
):
1481 with m
.If(stbs_done
& (adjust_acks
== 1)):
1482 sync
+= r1
.state
.eq(State
.IDLE
)
1483 sync
+= r1
.wb
.cyc
.eq(0)
1484 sync
+= r1
.wb
.stb
.eq(0)
1485 sync
+= r1
.dec_acks
.eq(1)
1487 with m
.Case(State
.NC_LOAD_WAIT_ACK
):
1488 # Clear stb when slave accepted request
1489 with m
.If(~wb_in
.stall
):
1490 sync
+= r1
.wb
.stb
.eq(0)
1492 # Got ack ? complete.
1493 with m
.If(wb_in
.ack
):
1494 sync
+= r1
.state
.eq(State
.IDLE
)
1495 sync
+= r1
.full
.eq(0)
1496 sync
+= r1
.slow_valid
.eq(1)
1498 with m
.If(~r1
.mmu_req
):
1499 sync
+= r1
.ls_valid
.eq(1)
1501 sync
+= r1
.mmu_done
.eq(1)
1503 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1504 sync
+= r1
.use_forward1
.eq(1)
1505 sync
+= r1
.wb
.cyc
.eq(0)
1506 sync
+= r1
.wb
.stb
.eq(0)
1508 def dcache_log(self
, m
, r1
, valid_ra
, tlb_hit_way
, stall_out
):
1511 d_out
, wb_in
, log_out
= self
.d_out
, self
.wb_in
, self
.log_out
1513 sync
+= log_out
.eq(Cat(r1
.state
[:3], valid_ra
, tlb_hit_way
[:3],
1514 stall_out
, req_op
[:3], d_out
.valid
, d_out
.error
,
1515 r1
.wb
.cyc
, r1
.wb
.stb
, wb_in
.ack
, wb_in
.stall
,
1518 def elaborate(self
, platform
):
1523 # Storage. Hopefully "cache_rows" is a BRAM, the rest is LUTs
1524 cache_tags
= CacheTagArray()
1525 cache_tag_set
= Signal(TAG_RAM_WIDTH
)
1526 cache_valid_bits
= CacheValidBitsArray()
1528 # TODO attribute ram_style : string;
1529 # TODO attribute ram_style of cache_tags : signal is "distributed";
1531 """note: these are passed to nmigen.hdl.Memory as "attributes".
1532 don't know how, just that they are.
1534 dtlb_valid_bits
= TLBValidBitsArray()
1535 dtlb_tags
= TLBTagsArray()
1536 dtlb_ptes
= TLBPtesArray()
1537 # TODO attribute ram_style of
1538 # dtlb_tags : signal is "distributed";
1539 # TODO attribute ram_style of
1540 # dtlb_ptes : signal is "distributed";
1542 r0
= RegStage0("r0")
1545 r1
= RegStage1("r1")
1547 reservation
= Reservation()
1549 # Async signals on incoming request
1550 req_index
= Signal(INDEX_BITS
)
1551 req_row
= Signal(ROW_BITS
)
1552 req_hit_way
= Signal(WAY_BITS
)
1553 req_tag
= Signal(TAG_BITS
)
1555 req_data
= Signal(64)
1556 req_same_tag
= Signal()
1559 early_req_row
= Signal(ROW_BITS
)
1561 cancel_store
= Signal()
1563 clear_rsrv
= Signal()
1568 use_forward1_next
= Signal()
1569 use_forward2_next
= Signal()
1571 cache_out
= CacheRamOut()
1573 plru_victim
= PLRUOut()
1574 replace_way
= Signal(WAY_BITS
)
1576 # Wishbone read/write/cache write formatting signals
1580 tlb_tag_way
= Signal(TLB_TAG_WAY_BITS
)
1581 tlb_pte_way
= Signal(TLB_PTE_WAY_BITS
)
1582 tlb_valid_way
= Signal(TLB_NUM_WAYS
)
1583 tlb_req_index
= Signal(TLB_SET_BITS
)
1585 tlb_hit_way
= Signal(TLB_WAY_BITS
)
1586 pte
= Signal(TLB_PTE_BITS
)
1587 ra
= Signal(REAL_ADDR_BITS
)
1589 perm_attr
= PermAttr("dc_perms")
1592 access_ok
= Signal()
1594 tlb_plru_victim
= TLBPLRUOut()
1596 # we don't yet handle collisions between loadstore1 requests
1598 comb
+= self
.m_out
.stall
.eq(0)
1600 # Hold off the request in r0 when r1 has an uncompleted request
1601 comb
+= r0_stall
.eq(r0_full
& r1
.full
)
1602 comb
+= r0_valid
.eq(r0_full
& ~r1
.full
)
1603 comb
+= self
.stall_out
.eq(r0_stall
)
1605 # Wire up wishbone request latch out of stage 1
1606 comb
+= self
.wb_out
.eq(r1
.wb
)
1608 # call sub-functions putting everything together, using shared
1609 # signals established above
1610 self
.stage_0(m
, r0
, r1
, r0_full
)
1611 self
.tlb_read(m
, r0_stall
, tlb_valid_way
,
1612 tlb_tag_way
, tlb_pte_way
, dtlb_valid_bits
,
1613 dtlb_tags
, dtlb_ptes
)
1614 self
.tlb_search(m
, tlb_req_index
, r0
, r0_valid
,
1615 tlb_valid_way
, tlb_tag_way
, tlb_hit_way
,
1616 tlb_pte_way
, pte
, tlb_hit
, valid_ra
, perm_attr
, ra
)
1617 self
.tlb_update(m
, r0_valid
, r0
, dtlb_valid_bits
, tlb_req_index
,
1618 tlb_hit_way
, tlb_hit
, tlb_plru_victim
, tlb_tag_way
,
1619 dtlb_tags
, tlb_pte_way
, dtlb_ptes
)
1620 self
.maybe_plrus(m
, r1
, plru_victim
)
1621 self
.maybe_tlb_plrus(m
, r1
, tlb_plru_victim
)
1622 self
.cache_tag_read(m
, r0_stall
, req_index
, cache_tag_set
, cache_tags
)
1623 self
.dcache_request(m
, r0
, ra
, req_index
, req_row
, req_tag
,
1624 r0_valid
, r1
, cache_valid_bits
, replace_way
,
1625 use_forward1_next
, use_forward2_next
,
1626 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
1627 valid_ra
, perm_ok
, access_ok
, req_op
, req_go
,
1629 tlb_hit
, tlb_hit_way
, tlb_valid_way
, cache_tag_set
,
1630 cancel_store
, req_same_tag
, r0_stall
, early_req_row
)
1631 self
.reservation_comb(m
, cancel_store
, set_rsrv
, clear_rsrv
,
1632 r0_valid
, r0
, reservation
)
1633 self
.reservation_reg(m
, r0_valid
, access_ok
, set_rsrv
, clear_rsrv
,
1635 self
.writeback_control(m
, r1
, cache_out
)
1636 self
.rams(m
, r1
, early_req_row
, cache_out
, replace_way
)
1637 self
.dcache_fast_hit(m
, req_op
, r0_valid
, r0
, r1
,
1638 req_hit_way
, req_index
, req_tag
, access_ok
,
1639 tlb_hit
, tlb_hit_way
, tlb_req_index
)
1640 self
.dcache_slow(m
, r1
, use_forward1_next
, use_forward2_next
,
1641 cache_valid_bits
, r0
, replace_way
,
1642 req_hit_way
, req_same_tag
,
1643 r0_valid
, req_op
, cache_tags
, req_go
, ra
)
1644 #self.dcache_log(m, r1, valid_ra, tlb_hit_way, stall_out)
1648 def dcache_load(dut
, addr
, nc
=0):
1649 yield dut
.d_in
.load
.eq(1)
1650 yield dut
.d_in
.nc
.eq(nc
)
1651 yield dut
.d_in
.addr
.eq(addr
)
1652 yield dut
.d_in
.valid
.eq(1)
1654 yield dut
.d_in
.valid
.eq(0)
1656 while not (yield dut
.d_out
.valid
):
1658 data
= yield dut
.d_out
.data
1662 def dcache_store(dut
, addr
, data
, nc
=0):
1663 yield dut
.d_in
.load
.eq(0)
1664 yield dut
.d_in
.nc
.eq(nc
)
1665 yield dut
.d_in
.data
.eq(data
)
1666 yield dut
.d_in
.addr
.eq(addr
)
1667 yield dut
.d_in
.valid
.eq(1)
1669 yield dut
.d_in
.valid
.eq(0)
1671 while not (yield dut
.d_out
.valid
):
1675 def dcache_sim(dut
):
1677 yield dut
.d_in
.valid
.eq(0)
1678 yield dut
.d_in
.load
.eq(0)
1679 yield dut
.d_in
.priv_mode
.eq(1)
1680 yield dut
.d_in
.nc
.eq(0)
1681 yield dut
.d_in
.addr
.eq(0)
1682 yield dut
.d_in
.data
.eq(0)
1683 yield dut
.m_in
.valid
.eq(0)
1684 yield dut
.m_in
.addr
.eq(0)
1685 yield dut
.m_in
.pte
.eq(0)
1686 # wait 4 * clk_period
1692 # Cacheable read of address 4
1693 data
= yield from dcache_load(dut
, 0x4)
1694 addr
= yield dut
.d_in
.addr
1695 assert data
== 0x0000000100000000, \
1696 f
"data @%x=%x expected 0x0000000100000000" % (addr
, data
)
1698 # Cacheable read of address 30
1699 data
= yield from dcache_load(dut
, 0x30)
1700 addr
= yield dut
.d_in
.addr
1701 assert data
== 0x0000000D0000000C, \
1702 f
"data @%x=%x expected 0000000D0000000C" % (addr
, data
)
1704 # 2nd Cacheable read of address 30
1705 data
= yield from dcache_load(dut
, 0x30)
1706 addr
= yield dut
.d_in
.addr
1707 assert data
== 0x0000000D0000000C, \
1708 f
"data @%x=%x expected 0000000D0000000C" % (addr
, data
)
1710 # Non-cacheable read of address 100
1711 data
= yield from dcache_load(dut
, 0x100, nc
=1)
1712 addr
= yield dut
.d_in
.addr
1713 assert data
== 0x0000004100000040, \
1714 f
"data @%x=%x expected 0000004100000040" % (addr
, data
)
1716 # Store at address 30
1717 yield from dcache_store(dut
, 0x30, 0x12345678)
1728 # 3nd Cacheable read of address 30
1729 data
= yield from dcache_load(dut
, 0x30)
1730 addr
= yield dut
.d_in
.addr
1731 assert data
== 0x12345678, \
1732 f
"data @%x=%x expected 0x12345678" % (addr
, data
)
1742 vl
= rtlil
.convert(dut
, ports
=[])
1743 with
open("test_dcache.il", "w") as f
:
1747 for i
in range(0,128):
1748 mem
.append((i
*2)|
((i
*2+1)<<32))
1749 memory
= Memory(width
=64, depth
=16*8, init
=mem
)
1750 sram
= SRAM(memory
=memory
, granularity
=8)
1753 m
.submodules
.dcache
= dut
1754 m
.submodules
.sram
= sram
1756 m
.d
.comb
+= sram
.bus
.cyc
.eq(dut
.wb_out
.cyc
)
1757 m
.d
.comb
+= sram
.bus
.stb
.eq(dut
.wb_out
.stb
)
1758 m
.d
.comb
+= sram
.bus
.we
.eq(dut
.wb_out
.we
)
1759 m
.d
.comb
+= sram
.bus
.sel
.eq(dut
.wb_out
.sel
)
1760 m
.d
.comb
+= sram
.bus
.adr
.eq(dut
.wb_out
.adr
[3:])
1761 m
.d
.comb
+= sram
.bus
.dat_w
.eq(dut
.wb_out
.dat
)
1763 m
.d
.comb
+= dut
.wb_in
.ack
.eq(sram
.bus
.ack
)
1764 m
.d
.comb
+= dut
.wb_in
.dat
.eq(sram
.bus
.dat_r
)
1770 sim
.add_sync_process(wrap(dcache_sim(dut
)))
1771 with sim
.write_vcd('test_dcache.vcd'):
1774 if __name__
== '__main__':