3 based on Anton Blanchard microwatt dcache.vhdl
5 note that the microwatt dcache wishbone interface expects "stall".
6 for simplicity at the moment this is hard-coded to cyc & ~ack.
7 see WB4 spec, p84, section 5.2.1
11 sys
.setrecursionlimit(1000000)
13 from enum
import Enum
, unique
15 from nmigen
import Module
, Signal
, Elaboratable
, Cat
, Repl
, Array
, Const
16 from nmutil
.util
import Display
18 from copy
import deepcopy
19 from random
import randint
, seed
21 from nmigen
.cli
import main
22 from nmutil
.iocontrol
import RecordObject
23 from nmigen
.utils
import log2_int
24 from soc
.experiment
.mem_types
import (LoadStore1ToDCacheType
,
25 DCacheToLoadStore1Type
,
29 from soc
.experiment
.wb_types
import (WB_ADDR_BITS
, WB_DATA_BITS
, WB_SEL_BITS
,
30 WBAddrType
, WBDataType
, WBSelType
,
31 WBMasterOut
, WBSlaveOut
,
32 WBMasterOutVector
, WBSlaveOutVector
,
33 WBIOMasterOut
, WBIOSlaveOut
)
35 from soc
.experiment
.cache_ram
import CacheRam
36 #from soc.experiment.plru import PLRU
37 from nmutil
.plru
import PLRU
40 from soc
.bus
.sram
import SRAM
41 from nmigen
import Memory
42 from nmigen
.cli
import rtlil
44 # NOTE: to use cxxsim, export NMIGEN_SIM_MODE=cxxsim from the shell
45 # Also, check out the cxxsim nmigen branch, and latest yosys from git
46 from nmutil
.sim_tmp_alternative
import Simulator
48 from nmutil
.util
import wrap
51 # TODO: make these parameters of DCache at some point
52 LINE_SIZE
= 64 # Line size in bytes
53 NUM_LINES
= 16 # Number of lines in a set
54 NUM_WAYS
= 4 # Number of ways
55 TLB_SET_SIZE
= 64 # L1 DTLB entries per set
56 TLB_NUM_WAYS
= 2 # L1 DTLB number of sets
57 TLB_LG_PGSZ
= 12 # L1 DTLB log_2(page_size)
58 LOG_LENGTH
= 0 # Non-zero to enable log data collection
60 # BRAM organisation: We never access more than
61 # -- WB_DATA_BITS at a time so to save
62 # -- resources we make the array only that wide, and
63 # -- use consecutive indices for to make a cache "line"
65 # -- ROW_SIZE is the width in bytes of the BRAM
66 # -- (based on WB, so 64-bits)
67 ROW_SIZE
= WB_DATA_BITS
// 8;
69 # ROW_PER_LINE is the number of row (wishbone
70 # transactions) in a line
71 ROW_PER_LINE
= LINE_SIZE
// ROW_SIZE
73 # BRAM_ROWS is the number of rows in BRAM needed
74 # to represent the full dcache
75 BRAM_ROWS
= NUM_LINES
* ROW_PER_LINE
77 print ("ROW_SIZE", ROW_SIZE
)
78 print ("ROW_PER_LINE", ROW_PER_LINE
)
79 print ("BRAM_ROWS", BRAM_ROWS
)
80 print ("NUM_WAYS", NUM_WAYS
)
82 # Bit fields counts in the address
84 # REAL_ADDR_BITS is the number of real address
88 # ROW_BITS is the number of bits to select a row
89 ROW_BITS
= log2_int(BRAM_ROWS
)
91 # ROW_LINE_BITS is the number of bits to select
93 ROW_LINE_BITS
= log2_int(ROW_PER_LINE
)
95 # LINE_OFF_BITS is the number of bits for
96 # the offset in a cache line
97 LINE_OFF_BITS
= log2_int(LINE_SIZE
)
99 # ROW_OFF_BITS is the number of bits for
100 # the offset in a row
101 ROW_OFF_BITS
= log2_int(ROW_SIZE
)
103 # INDEX_BITS is the number if bits to
104 # select a cache line
105 INDEX_BITS
= log2_int(NUM_LINES
)
107 # SET_SIZE_BITS is the log base 2 of the set size
108 SET_SIZE_BITS
= LINE_OFF_BITS
+ INDEX_BITS
110 # TAG_BITS is the number of bits of
111 # the tag part of the address
112 TAG_BITS
= REAL_ADDR_BITS
- SET_SIZE_BITS
114 # TAG_WIDTH is the width in bits of each way of the tag RAM
115 TAG_WIDTH
= TAG_BITS
+ 7 - ((TAG_BITS
+ 7) % 8)
117 # WAY_BITS is the number of bits to select a way
118 WAY_BITS
= log2_int(NUM_WAYS
)
120 # Example of layout for 32 lines of 64 bytes:
122 .. tag |index| line |
124 .. | |---| | ROW_LINE_BITS (3)
125 .. | |--- - --| LINE_OFF_BITS (6)
126 .. | |- --| ROW_OFF_BITS (3)
127 .. |----- ---| | ROW_BITS (8)
128 .. |-----| | INDEX_BITS (5)
129 .. --------| | TAG_BITS (45)
132 print ("Dcache TAG %d IDX %d ROW_BITS %d ROFF %d LOFF %d RLB %d" % \
133 (TAG_BITS
, INDEX_BITS
, ROW_BITS
,
134 ROW_OFF_BITS
, LINE_OFF_BITS
, ROW_LINE_BITS
))
135 print ("index @: %d-%d" % (LINE_OFF_BITS
, SET_SIZE_BITS
))
136 print ("row @: %d-%d" % (LINE_OFF_BITS
, ROW_OFF_BITS
))
137 print ("tag @: %d-%d width %d" % (SET_SIZE_BITS
, REAL_ADDR_BITS
, TAG_WIDTH
))
139 TAG_RAM_WIDTH
= TAG_WIDTH
* NUM_WAYS
141 print ("TAG_RAM_WIDTH", TAG_RAM_WIDTH
)
144 return Array(Signal(TAG_RAM_WIDTH
, name
="cachetag_%d" % x
) \
145 for x
in range(NUM_LINES
))
147 def CacheValidBitsArray():
148 return Array(Signal(NUM_WAYS
, name
="cachevalid_%d" % x
) \
149 for x
in range(NUM_LINES
))
151 def RowPerLineValidArray():
152 return Array(Signal(name
="rows_valid%d" % x
) \
153 for x
in range(ROW_PER_LINE
))
156 TLB_SET_BITS
= log2_int(TLB_SET_SIZE
)
157 TLB_WAY_BITS
= log2_int(TLB_NUM_WAYS
)
158 TLB_EA_TAG_BITS
= 64 - (TLB_LG_PGSZ
+ TLB_SET_BITS
)
159 TLB_TAG_WAY_BITS
= TLB_NUM_WAYS
* TLB_EA_TAG_BITS
161 TLB_PTE_WAY_BITS
= TLB_NUM_WAYS
* TLB_PTE_BITS
;
164 return (1<<log2_int(x
, False)) == x
166 assert (LINE_SIZE
% ROW_SIZE
) == 0, "LINE_SIZE not multiple of ROW_SIZE"
167 assert ispow2(LINE_SIZE
), "LINE_SIZE not power of 2"
168 assert ispow2(NUM_LINES
), "NUM_LINES not power of 2"
169 assert ispow2(ROW_PER_LINE
), "ROW_PER_LINE not power of 2"
170 assert ROW_BITS
== (INDEX_BITS
+ ROW_LINE_BITS
), "geometry bits don't add up"
171 assert (LINE_OFF_BITS
== ROW_OFF_BITS
+ ROW_LINE_BITS
), \
172 "geometry bits don't add up"
173 assert REAL_ADDR_BITS
== (TAG_BITS
+ INDEX_BITS
+ LINE_OFF_BITS
), \
174 "geometry bits don't add up"
175 assert REAL_ADDR_BITS
== (TAG_BITS
+ ROW_BITS
+ ROW_OFF_BITS
), \
176 "geometry bits don't add up"
177 assert 64 == WB_DATA_BITS
, "Can't yet handle wb width that isn't 64-bits"
178 assert SET_SIZE_BITS
<= TLB_LG_PGSZ
, "Set indexed by virtual address"
181 def TLBValidBitsArray():
182 return Array(Signal(TLB_NUM_WAYS
, name
="tlbvalid%d" % x
) \
183 for x
in range(TLB_SET_SIZE
))
186 return Array(Signal(TLB_EA_TAG_BITS
, name
="tlbtagea%d" % x
) \
187 for x
in range (TLB_NUM_WAYS
))
190 return Array(Signal(TLB_TAG_WAY_BITS
, name
="tlbtags%d" % x
) \
191 for x
in range (TLB_SET_SIZE
))
194 return Array(Signal(TLB_PTE_WAY_BITS
, name
="tlbptes%d" % x
) \
195 for x
in range(TLB_SET_SIZE
))
198 return Array(Signal(WAY_BITS
, name
="hitway_%d" % x
) \
199 for x
in range(TLB_NUM_WAYS
))
201 # Cache RAM interface
203 return Array(Signal(WB_DATA_BITS
, name
="cache_out%d" % x
) \
204 for x
in range(NUM_WAYS
))
206 # PLRU output interface
208 return Array(Signal(WAY_BITS
, name
="plru_out%d" % x
) \
209 for x
in range(NUM_LINES
))
211 # TLB PLRU output interface
213 return Array(Signal(TLB_WAY_BITS
, name
="tlbplru_out%d" % x
) \
214 for x
in range(TLB_SET_SIZE
))
216 # Helper functions to decode incoming requests
218 # Return the cache line index (tag index) for an address
220 return addr
[LINE_OFF_BITS
:SET_SIZE_BITS
]
222 # Return the cache row index (data memory) for an address
224 return addr
[ROW_OFF_BITS
:SET_SIZE_BITS
]
226 # Return the index of a row within a line
227 def get_row_of_line(row
):
228 return row
[:ROW_BITS
][:ROW_LINE_BITS
]
230 # Returns whether this is the last row of a line
231 def is_last_row_addr(addr
, last
):
232 return addr
[ROW_OFF_BITS
:LINE_OFF_BITS
] == last
234 # Returns whether this is the last row of a line
235 def is_last_row(row
, last
):
236 return get_row_of_line(row
) == last
238 # Return the next row in the current cache line. We use a
239 # dedicated function in order to limit the size of the
240 # generated adder to be only the bits within a cache line
241 # (3 bits with default settings)
243 row_v
= row
[0:ROW_LINE_BITS
] + 1
244 return Cat(row_v
[:ROW_LINE_BITS
], row
[ROW_LINE_BITS
:])
246 # Get the tag value from the address
248 return addr
[SET_SIZE_BITS
:REAL_ADDR_BITS
]
250 # Read a tag from a tag memory row
251 def read_tag(way
, tagset
):
252 return tagset
.word_select(way
, TAG_WIDTH
)[:TAG_BITS
]
254 # Read a TLB tag from a TLB tag memory row
255 def read_tlb_tag(way
, tags
):
256 return tags
.word_select(way
, TLB_EA_TAG_BITS
)
258 # Write a TLB tag to a TLB tag memory row
259 def write_tlb_tag(way
, tags
, tag
):
260 return read_tlb_tag(way
, tags
).eq(tag
)
262 # Read a PTE from a TLB PTE memory row
263 def read_tlb_pte(way
, ptes
):
264 return ptes
.word_select(way
, TLB_PTE_BITS
)
266 def write_tlb_pte(way
, ptes
, newpte
):
267 return read_tlb_pte(way
, ptes
).eq(newpte
)
270 # Record for storing permission, attribute, etc. bits from a PTE
271 class PermAttr(RecordObject
):
272 def __init__(self
, name
=None):
273 super().__init
__(name
=name
)
274 self
.reference
= Signal()
275 self
.changed
= Signal()
276 self
.nocache
= Signal()
278 self
.rd_perm
= Signal()
279 self
.wr_perm
= Signal()
282 def extract_perm_attr(pte
):
287 # Type of operation on a "valid" input
291 OP_BAD
= 1 # NC cache hit, TLB miss, prot/RC failure
292 OP_STCX_FAIL
= 2 # conditional store w/o reservation
293 OP_LOAD_HIT
= 3 # Cache hit on load
294 OP_LOAD_MISS
= 4 # Load missing cache
295 OP_LOAD_NC
= 5 # Non-cachable load
296 OP_STORE_HIT
= 6 # Store hitting cache
297 OP_STORE_MISS
= 7 # Store missing cache
300 # Cache state machine
303 IDLE
= 0 # Normal load hit processing
304 RELOAD_WAIT_ACK
= 1 # Cache reload wait ack
305 STORE_WAIT_ACK
= 2 # Store wait ack
306 NC_LOAD_WAIT_ACK
= 3 # Non-cachable load wait ack
311 # In order to make timing, we use the BRAMs with
312 # an output buffer, which means that the BRAM
313 # output is delayed by an extra cycle.
315 # Thus, the dcache has a 2-stage internal pipeline
316 # for cache hits with no stalls.
318 # All other operations are handled via stalling
319 # in the first stage.
321 # The second stage can thus complete a hit at the same
322 # time as the first stage emits a stall for a complex op.
324 # Stage 0 register, basically contains just the latched request
326 class RegStage0(RecordObject
):
327 def __init__(self
, name
=None):
328 super().__init
__(name
=name
)
329 self
.req
= LoadStore1ToDCacheType(name
="lsmem")
330 self
.tlbie
= Signal()
331 self
.doall
= Signal()
332 self
.tlbld
= Signal()
333 self
.mmu_req
= Signal() # indicates source of request
336 class MemAccessRequest(RecordObject
):
337 def __init__(self
, name
=None):
338 super().__init
__(name
=name
)
340 self
.valid
= Signal()
342 self
.real_addr
= Signal(REAL_ADDR_BITS
)
343 self
.data
= Signal(64)
344 self
.byte_sel
= Signal(8)
345 self
.hit_way
= Signal(WAY_BITS
)
346 self
.same_tag
= Signal()
347 self
.mmu_req
= Signal()
350 # First stage register, contains state for stage 1 of load hits
351 # and for the state machine used by all other operations
352 class RegStage1(RecordObject
):
353 def __init__(self
, name
=None):
354 super().__init
__(name
=name
)
355 # Info about the request
356 self
.full
= Signal() # have uncompleted request
357 self
.mmu_req
= Signal() # request is from MMU
358 self
.req
= MemAccessRequest(name
="reqmem")
361 self
.hit_way
= Signal(WAY_BITS
)
362 self
.hit_load_valid
= Signal()
363 self
.hit_index
= Signal(INDEX_BITS
)
364 self
.cache_hit
= Signal()
367 self
.tlb_hit
= Signal()
368 self
.tlb_hit_way
= Signal(TLB_NUM_WAYS
)
369 self
.tlb_hit_index
= Signal(TLB_WAY_BITS
)
371 # 2-stage data buffer for data forwarded from writes to reads
372 self
.forward_data1
= Signal(64)
373 self
.forward_data2
= Signal(64)
374 self
.forward_sel1
= Signal(8)
375 self
.forward_valid1
= Signal()
376 self
.forward_way1
= Signal(WAY_BITS
)
377 self
.forward_row1
= Signal(ROW_BITS
)
378 self
.use_forward1
= Signal()
379 self
.forward_sel
= Signal(8)
381 # Cache miss state (reload state machine)
382 self
.state
= Signal(State
)
384 self
.write_bram
= Signal()
385 self
.write_tag
= Signal()
386 self
.slow_valid
= Signal()
387 self
.real_adr
= Signal(REAL_ADDR_BITS
)
388 self
.wb
= WBMasterOut("wb")
389 self
.reload_tag
= Signal(TAG_BITS
)
390 self
.store_way
= Signal(WAY_BITS
)
391 self
.store_row
= Signal(ROW_BITS
)
392 self
.store_index
= Signal(INDEX_BITS
)
393 self
.end_row_ix
= Signal(ROW_LINE_BITS
)
394 self
.rows_valid
= RowPerLineValidArray()
395 self
.acks_pending
= Signal(3)
396 self
.inc_acks
= Signal()
397 self
.dec_acks
= Signal()
399 # Signals to complete (possibly with error)
400 self
.ls_valid
= Signal()
401 self
.ls_error
= Signal()
402 self
.mmu_done
= Signal()
403 self
.mmu_error
= Signal()
404 self
.cache_paradox
= Signal()
406 # Signal to complete a failed stcx.
407 self
.stcx_fail
= Signal()
410 # Reservation information
411 class Reservation(RecordObject
):
414 self
.valid
= Signal()
415 self
.addr
= Signal(64-LINE_OFF_BITS
)
418 class DTLBUpdate(Elaboratable
):
420 self
.tlbie
= Signal()
421 self
.tlbwe
= Signal()
422 self
.doall
= Signal()
423 self
.updated
= Signal()
424 self
.v_updated
= Signal()
425 self
.tlb_hit
= Signal()
426 self
.tlb_req_index
= Signal(TLB_SET_BITS
)
428 self
.tlb_hit_way
= Signal(TLB_WAY_BITS
)
429 self
.tlb_tag_way
= Signal(TLB_TAG_WAY_BITS
)
430 self
.tlb_pte_way
= Signal(TLB_PTE_WAY_BITS
)
431 self
.repl_way
= Signal(TLB_WAY_BITS
)
432 self
.eatag
= Signal(TLB_EA_TAG_BITS
)
433 self
.pte_data
= Signal(TLB_PTE_BITS
)
435 self
.dv
= Signal(TLB_NUM_WAYS
) # tlb_way_valids_t
437 self
.tb_out
= Signal(TLB_TAG_WAY_BITS
) # tlb_way_tags_t
438 self
.pb_out
= Signal(TLB_NUM_WAYS
) # tlb_way_valids_t
439 self
.db_out
= Signal(TLB_PTE_WAY_BITS
) # tlb_way_ptes_t
441 def elaborate(self
, platform
):
446 tagset
= Signal(TLB_TAG_WAY_BITS
)
447 pteset
= Signal(TLB_PTE_WAY_BITS
)
449 tb_out
, pb_out
, db_out
= self
.tb_out
, self
.pb_out
, self
.db_out
450 comb
+= db_out
.eq(self
.dv
)
452 with m
.If(self
.tlbie
& self
.doall
):
453 pass # clear all back in parent
454 with m
.Elif(self
.tlbie
):
455 with m
.If(self
.tlb_hit
):
456 comb
+= db_out
.bit_select(self
.tlb_hit_way
, 1).eq(1)
457 comb
+= self
.v_updated
.eq(1)
459 with m
.Elif(self
.tlbwe
):
461 comb
+= tagset
.eq(self
.tlb_tag_way
)
462 comb
+= write_tlb_tag(self
.repl_way
, tagset
, self
.eatag
)
463 comb
+= tb_out
.eq(tagset
)
465 comb
+= pteset
.eq(self
.tlb_pte_way
)
466 comb
+= write_tlb_pte(self
.repl_way
, pteset
, self
.pte_data
)
467 comb
+= pb_out
.eq(pteset
)
469 comb
+= db_out
.bit_select(self
.repl_way
, 1).eq(1)
471 comb
+= self
.updated
.eq(1)
472 comb
+= self
.v_updated
.eq(1)
477 class DCachePendingHit(Elaboratable
):
479 def __init__(self
, tlb_pte_way
, tlb_valid_way
, tlb_hit_way
,
480 cache_valid_idx
, cache_tag_set
,
485 self
.virt_mode
= Signal()
486 self
.is_hit
= Signal()
487 self
.tlb_hit
= Signal()
488 self
.hit_way
= Signal(WAY_BITS
)
489 self
.rel_match
= Signal()
490 self
.req_index
= Signal(INDEX_BITS
)
491 self
.reload_tag
= Signal(TAG_BITS
)
493 self
.tlb_hit_way
= tlb_hit_way
494 self
.tlb_pte_way
= tlb_pte_way
495 self
.tlb_valid_way
= tlb_valid_way
496 self
.cache_valid_idx
= cache_valid_idx
497 self
.cache_tag_set
= cache_tag_set
498 self
.req_addr
= req_addr
499 self
.hit_set
= hit_set
501 def elaborate(self
, platform
):
507 virt_mode
= self
.virt_mode
509 tlb_pte_way
= self
.tlb_pte_way
510 tlb_valid_way
= self
.tlb_valid_way
511 cache_valid_idx
= self
.cache_valid_idx
512 cache_tag_set
= self
.cache_tag_set
513 req_addr
= self
.req_addr
514 tlb_hit_way
= self
.tlb_hit_way
515 tlb_hit
= self
.tlb_hit
516 hit_set
= self
.hit_set
517 hit_way
= self
.hit_way
518 rel_match
= self
.rel_match
519 req_index
= self
.req_index
520 reload_tag
= self
.reload_tag
522 rel_matches
= Array(Signal(name
="rel_matches_%d" % i
) \
523 for i
in range(TLB_NUM_WAYS
))
524 hit_way_set
= HitWaySet()
526 # Test if pending request is a hit on any way
527 # In order to make timing in virtual mode,
528 # when we are using the TLB, we compare each
529 # way with each of the real addresses from each way of
530 # the TLB, and then decide later which match to use.
532 with m
.If(virt_mode
):
533 for j
in range(TLB_NUM_WAYS
): # tlb_num_way_t
534 s_tag
= Signal(TAG_BITS
, name
="s_tag%d" % j
)
536 s_pte
= Signal(TLB_PTE_BITS
)
537 s_ra
= Signal(REAL_ADDR_BITS
)
538 comb
+= s_pte
.eq(read_tlb_pte(j
, tlb_pte_way
))
539 comb
+= s_ra
.eq(Cat(req_addr
[0:TLB_LG_PGSZ
],
540 s_pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
541 comb
+= s_tag
.eq(get_tag(s_ra
))
543 for i
in range(NUM_WAYS
): # way_t
544 is_tag_hit
= Signal(name
="is_tag_hit_%d_%d" % (j
, i
))
545 comb
+= is_tag_hit
.eq(go
& cache_valid_idx
[i
] &
546 (read_tag(i
, cache_tag_set
) == s_tag
)
548 with m
.If(is_tag_hit
):
549 comb
+= hit_way_set
[j
].eq(i
)
551 comb
+= hit_set
[j
].eq(s_hit
)
552 with m
.If(s_tag
== reload_tag
):
553 comb
+= rel_matches
[j
].eq(1)
555 comb
+= is_hit
.eq(hit_set
[tlb_hit_way
])
556 comb
+= hit_way
.eq(hit_way_set
[tlb_hit_way
])
557 comb
+= rel_match
.eq(rel_matches
[tlb_hit_way
])
559 s_tag
= Signal(TAG_BITS
)
560 comb
+= s_tag
.eq(get_tag(req_addr
))
561 for i
in range(NUM_WAYS
): # way_t
562 is_tag_hit
= Signal(name
="is_tag_hit_%d" % i
)
563 comb
+= is_tag_hit
.eq(go
& cache_valid_idx
[i
] &
564 (read_tag(i
, cache_tag_set
) == s_tag
))
565 with m
.If(is_tag_hit
):
566 comb
+= hit_way
.eq(i
)
568 with m
.If(s_tag
== reload_tag
):
569 comb
+= rel_match
.eq(1)
574 class DCache(Elaboratable
):
575 """Set associative dcache write-through
576 TODO (in no specific order):
577 * See list in icache.vhdl
578 * Complete load misses on the cycle when WB data comes instead of
579 at the end of line (this requires dealing with requests coming in
583 self
.d_in
= LoadStore1ToDCacheType("d_in")
584 self
.d_out
= DCacheToLoadStore1Type("d_out")
586 self
.m_in
= MMUToDCacheType("m_in")
587 self
.m_out
= DCacheToMMUType("m_out")
589 self
.stall_out
= Signal()
591 self
.wb_out
= WBMasterOut()
592 self
.wb_in
= WBSlaveOut()
594 self
.log_out
= Signal(20)
596 def stage_0(self
, m
, r0
, r1
, r0_full
):
597 """Latch the request in r0.req as long as we're not stalling
601 d_in
, d_out
, m_in
= self
.d_in
, self
.d_out
, self
.m_in
603 r
= RegStage0("stage0")
605 # TODO, this goes in unit tests and formal proofs
606 with m
.If(d_in
.valid
& m_in
.valid
):
607 sync
+= Display("request collision loadstore vs MMU")
609 with m
.If(m_in
.valid
):
610 comb
+= r
.req
.valid
.eq(1)
611 comb
+= r
.req
.load
.eq(~
(m_in
.tlbie | m_in
.tlbld
))
612 comb
+= r
.req
.dcbz
.eq(0)
613 comb
+= r
.req
.nc
.eq(0)
614 comb
+= r
.req
.reserve
.eq(0)
615 comb
+= r
.req
.virt_mode
.eq(0)
616 comb
+= r
.req
.priv_mode
.eq(1)
617 comb
+= r
.req
.addr
.eq(m_in
.addr
)
618 comb
+= r
.req
.data
.eq(m_in
.pte
)
619 comb
+= r
.req
.byte_sel
.eq(~
0) # Const -1 sets all to 0b111....
620 comb
+= r
.tlbie
.eq(m_in
.tlbie
)
621 comb
+= r
.doall
.eq(m_in
.doall
)
622 comb
+= r
.tlbld
.eq(m_in
.tlbld
)
623 comb
+= r
.mmu_req
.eq(1)
625 comb
+= r
.req
.eq(d_in
)
626 comb
+= r
.tlbie
.eq(0)
627 comb
+= r
.doall
.eq(0)
628 comb
+= r
.tlbld
.eq(0)
629 comb
+= r
.mmu_req
.eq(0)
630 with m
.If(~
(r1
.full
& r0_full
)):
632 sync
+= r0_full
.eq(r
.req
.valid
)
634 def tlb_read(self
, m
, r0_stall
, tlb_valid_way
,
635 tlb_tag_way
, tlb_pte_way
, dtlb_valid_bits
,
636 dtlb_tags
, dtlb_ptes
):
638 Operates in the second cycle on the request latched in r0.req.
639 TLB updates write the entry at the end of the second cycle.
643 m_in
, d_in
= self
.m_in
, self
.d_in
645 index
= Signal(TLB_SET_BITS
)
646 addrbits
= Signal(TLB_SET_BITS
)
649 amax
= TLB_LG_PGSZ
+ TLB_SET_BITS
651 with m
.If(m_in
.valid
):
652 comb
+= addrbits
.eq(m_in
.addr
[amin
: amax
])
654 comb
+= addrbits
.eq(d_in
.addr
[amin
: amax
])
655 comb
+= index
.eq(addrbits
)
657 # If we have any op and the previous op isn't finished,
658 # then keep the same output for next cycle.
659 with m
.If(~r0_stall
):
660 sync
+= tlb_valid_way
.eq(dtlb_valid_bits
[index
])
661 sync
+= tlb_tag_way
.eq(dtlb_tags
[index
])
662 sync
+= tlb_pte_way
.eq(dtlb_ptes
[index
])
664 def maybe_tlb_plrus(self
, m
, r1
, tlb_plru_victim
):
665 """Generate TLB PLRUs
670 if TLB_NUM_WAYS
== 0:
672 for i
in range(TLB_SET_SIZE
):
674 tlb_plru
= PLRU(TLB_WAY_BITS
)
675 setattr(m
.submodules
, "maybe_plru_%d" % i
, tlb_plru
)
676 tlb_plru_acc_en
= Signal()
678 comb
+= tlb_plru_acc_en
.eq(r1
.tlb_hit
& (r1
.tlb_hit_index
== i
))
679 comb
+= tlb_plru
.acc_en
.eq(tlb_plru_acc_en
)
680 comb
+= tlb_plru
.acc_i
.eq(r1
.tlb_hit_way
)
681 comb
+= tlb_plru_victim
[i
].eq(tlb_plru
.lru_o
)
683 def tlb_search(self
, m
, tlb_req_index
, r0
, r0_valid
,
684 tlb_valid_way
, tlb_tag_way
, tlb_hit_way
,
685 tlb_pte_way
, pte
, tlb_hit
, valid_ra
, perm_attr
, ra
):
689 hitway
= Signal(TLB_WAY_BITS
)
691 eatag
= Signal(TLB_EA_TAG_BITS
)
693 TLB_LG_END
= TLB_LG_PGSZ
+ TLB_SET_BITS
694 comb
+= tlb_req_index
.eq(r0
.req
.addr
[TLB_LG_PGSZ
: TLB_LG_END
])
695 comb
+= eatag
.eq(r0
.req
.addr
[TLB_LG_END
: 64 ])
697 for i
in range(TLB_NUM_WAYS
):
698 is_tag_hit
= Signal()
699 comb
+= is_tag_hit
.eq(tlb_valid_way
[i
]
700 & (read_tlb_tag(i
, tlb_tag_way
) == eatag
))
701 with m
.If(is_tag_hit
):
705 comb
+= tlb_hit
.eq(hit
& r0_valid
)
706 comb
+= tlb_hit_way
.eq(hitway
)
709 comb
+= pte
.eq(read_tlb_pte(hitway
, tlb_pte_way
))
710 comb
+= valid_ra
.eq(tlb_hit | ~r0
.req
.virt_mode
)
712 with m
.If(r0
.req
.virt_mode
):
713 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
714 r0
.req
.addr
[ROW_OFF_BITS
:TLB_LG_PGSZ
],
715 pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
716 comb
+= perm_attr
.reference
.eq(pte
[8])
717 comb
+= perm_attr
.changed
.eq(pte
[7])
718 comb
+= perm_attr
.nocache
.eq(pte
[5])
719 comb
+= perm_attr
.priv
.eq(pte
[3])
720 comb
+= perm_attr
.rd_perm
.eq(pte
[2])
721 comb
+= perm_attr
.wr_perm
.eq(pte
[1])
723 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
724 r0
.req
.addr
[ROW_OFF_BITS
:REAL_ADDR_BITS
]))
725 comb
+= perm_attr
.reference
.eq(1)
726 comb
+= perm_attr
.changed
.eq(1)
727 comb
+= perm_attr
.nocache
.eq(0)
728 comb
+= perm_attr
.priv
.eq(1)
729 comb
+= perm_attr
.rd_perm
.eq(1)
730 comb
+= perm_attr
.wr_perm
.eq(1)
732 def tlb_update(self
, m
, r0_valid
, r0
, dtlb_valid_bits
, tlb_req_index
,
733 tlb_hit_way
, tlb_hit
, tlb_plru_victim
, tlb_tag_way
,
734 dtlb_tags
, tlb_pte_way
, dtlb_ptes
):
736 dtlb_valids
= TLBValidBitsArray()
744 comb
+= tlbie
.eq(r0_valid
& r0
.tlbie
)
745 comb
+= tlbwe
.eq(r0_valid
& r0
.tlbld
)
747 m
.submodules
.tlb_update
= d
= DTLBUpdate()
748 with m
.If(tlbie
& r0
.doall
):
749 # clear all valid bits at once
750 for i
in range(TLB_SET_SIZE
):
751 sync
+= dtlb_valid_bits
[i
].eq(0)
752 with m
.If(d
.updated
):
753 sync
+= dtlb_tags
[tlb_req_index
].eq(d
.tb_out
)
754 sync
+= dtlb_ptes
[tlb_req_index
].eq(d
.pb_out
)
755 with m
.If(d
.v_updated
):
756 sync
+= dtlb_valid_bits
[tlb_req_index
].eq(d
.db_out
)
758 comb
+= d
.dv
.eq(dtlb_valid_bits
[tlb_req_index
])
760 comb
+= d
.tlbie
.eq(tlbie
)
761 comb
+= d
.tlbwe
.eq(tlbwe
)
762 comb
+= d
.doall
.eq(r0
.doall
)
763 comb
+= d
.tlb_hit
.eq(tlb_hit
)
764 comb
+= d
.tlb_hit_way
.eq(tlb_hit_way
)
765 comb
+= d
.tlb_tag_way
.eq(tlb_tag_way
)
766 comb
+= d
.tlb_pte_way
.eq(tlb_pte_way
)
767 comb
+= d
.tlb_req_index
.eq(tlb_req_index
)
770 comb
+= d
.repl_way
.eq(tlb_hit_way
)
772 comb
+= d
.repl_way
.eq(tlb_plru_victim
[tlb_req_index
])
773 comb
+= d
.eatag
.eq(r0
.req
.addr
[TLB_LG_PGSZ
+ TLB_SET_BITS
:64])
774 comb
+= d
.pte_data
.eq(r0
.req
.data
)
776 def maybe_plrus(self
, m
, r1
, plru_victim
):
782 if TLB_NUM_WAYS
== 0:
785 for i
in range(NUM_LINES
):
787 plru
= PLRU(WAY_BITS
)
788 setattr(m
.submodules
, "plru%d" % i
, plru
)
789 plru_acc_en
= Signal()
791 comb
+= plru_acc_en
.eq(r1
.cache_hit
& (r1
.hit_index
== i
))
792 comb
+= plru
.acc_en
.eq(plru_acc_en
)
793 comb
+= plru
.acc_i
.eq(r1
.hit_way
)
794 comb
+= plru_victim
[i
].eq(plru
.lru_o
)
796 def cache_tag_read(self
, m
, r0_stall
, req_index
, cache_tag_set
, cache_tags
):
797 """Cache tag RAM read port
801 m_in
, d_in
= self
.m_in
, self
.d_in
803 index
= Signal(INDEX_BITS
)
806 comb
+= index
.eq(req_index
)
807 with m
.Elif(m_in
.valid
):
808 comb
+= index
.eq(get_index(m_in
.addr
))
810 comb
+= index
.eq(get_index(d_in
.addr
))
811 sync
+= cache_tag_set
.eq(cache_tags
[index
])
813 def dcache_request(self
, m
, r0
, ra
, req_index
, req_row
, req_tag
,
814 r0_valid
, r1
, cache_valids
, replace_way
,
815 use_forward1_next
, use_forward2_next
,
816 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
817 valid_ra
, perm_ok
, access_ok
, req_op
, req_go
,
819 tlb_hit
, tlb_hit_way
, tlb_valid_way
, cache_tag_set
,
820 cancel_store
, req_same_tag
, r0_stall
, early_req_row
):
821 """Cache request parsing and hit detection
826 m_in
, d_in
= self
.m_in
, self
.d_in
829 hit_way
= Signal(WAY_BITS
)
834 hit_set
= Array(Signal(name
="hit_set_%d" % i
) \
835 for i
in range(TLB_NUM_WAYS
))
836 cache_valid_idx
= Signal(NUM_WAYS
)
838 # Extract line, row and tag from request
839 comb
+= req_index
.eq(get_index(r0
.req
.addr
))
840 comb
+= req_row
.eq(get_row(r0
.req
.addr
))
841 comb
+= req_tag
.eq(get_tag(ra
))
843 if False: # display on comb is a bit... busy.
844 comb
+= Display("dcache_req addr:%x ra: %x idx: %x tag: %x row: %x",
845 r0
.req
.addr
, ra
, req_index
, req_tag
, req_row
)
847 comb
+= go
.eq(r0_valid
& ~
(r0
.tlbie | r0
.tlbld
) & ~r1
.ls_error
)
848 comb
+= cache_valid_idx
.eq(cache_valids
[req_index
])
850 m
.submodules
.dcache_pend
= dc
= DCachePendingHit(tlb_pte_way
,
851 tlb_valid_way
, tlb_hit_way
,
852 cache_valid_idx
, cache_tag_set
,
856 comb
+= dc
.tlb_hit
.eq(tlb_hit
)
857 comb
+= dc
.reload_tag
.eq(r1
.reload_tag
)
858 comb
+= dc
.virt_mode
.eq(r0
.req
.virt_mode
)
860 comb
+= dc
.req_index
.eq(req_index
)
861 comb
+= is_hit
.eq(dc
.is_hit
)
862 comb
+= hit_way
.eq(dc
.hit_way
)
863 comb
+= req_same_tag
.eq(dc
.rel_match
)
865 # See if the request matches the line currently being reloaded
866 with m
.If((r1
.state
== State
.RELOAD_WAIT_ACK
) &
867 (req_index
== r1
.store_index
) & req_same_tag
):
868 # For a store, consider this a hit even if the row isn't
869 # valid since it will be by the time we perform the store.
870 # For a load, check the appropriate row valid bit.
871 rrow
= Signal(ROW_LINE_BITS
)
872 comb
+= rrow
.eq(req_row
)
873 valid
= r1
.rows_valid
[rrow
]
874 comb
+= is_hit
.eq((~r0
.req
.load
) | valid
)
875 comb
+= hit_way
.eq(replace_way
)
877 # Whether to use forwarded data for a load or not
878 with m
.If((get_row(r1
.req
.real_addr
) == req_row
) &
879 (r1
.req
.hit_way
== hit_way
)):
880 # Only need to consider r1.write_bram here, since if we
881 # are writing refill data here, then we don't have a
882 # cache hit this cycle on the line being refilled.
883 # (There is the possibility that the load following the
884 # load miss that started the refill could be to the old
885 # contents of the victim line, since it is a couple of
886 # cycles after the refill starts before we see the updated
887 # cache tag. In that case we don't use the bypass.)
888 comb
+= use_forward1_next
.eq(r1
.write_bram
)
889 with m
.If((r1
.forward_row1
== req_row
) & (r1
.forward_way1
== hit_way
)):
890 comb
+= use_forward2_next
.eq(r1
.forward_valid1
)
892 # The way that matched on a hit
893 comb
+= req_hit_way
.eq(hit_way
)
895 # The way to replace on a miss
896 with m
.If(r1
.write_tag
):
897 comb
+= replace_way
.eq(plru_victim
[r1
.store_index
])
899 comb
+= replace_way
.eq(r1
.store_way
)
901 # work out whether we have permission for this access
902 # NB we don't yet implement AMR, thus no KUAP
903 comb
+= rc_ok
.eq(perm_attr
.reference
904 & (r0
.req
.load | perm_attr
.changed
))
905 comb
+= perm_ok
.eq((r0
.req
.priv_mode |
(~perm_attr
.priv
)) &
907 (r0
.req
.load
& perm_attr
.rd_perm
)))
908 comb
+= access_ok
.eq(valid_ra
& perm_ok
& rc_ok
)
909 # Combine the request and cache hit status to decide what
910 # operation needs to be done
911 comb
+= nc
.eq(r0
.req
.nc | perm_attr
.nocache
)
912 comb
+= op
.eq(Op
.OP_NONE
)
914 with m
.If(~access_ok
):
915 comb
+= op
.eq(Op
.OP_BAD
)
916 with m
.Elif(cancel_store
):
917 comb
+= op
.eq(Op
.OP_STCX_FAIL
)
919 comb
+= opsel
.eq(Cat(is_hit
, nc
, r0
.req
.load
))
920 with m
.Switch(opsel
):
921 with m
.Case(0b101): comb
+= op
.eq(Op
.OP_LOAD_HIT
)
922 with m
.Case(0b100): comb
+= op
.eq(Op
.OP_LOAD_MISS
)
923 with m
.Case(0b110): comb
+= op
.eq(Op
.OP_LOAD_NC
)
924 with m
.Case(0b001): comb
+= op
.eq(Op
.OP_STORE_HIT
)
925 with m
.Case(0b000): comb
+= op
.eq(Op
.OP_STORE_MISS
)
926 with m
.Case(0b010): comb
+= op
.eq(Op
.OP_STORE_MISS
)
927 with m
.Case(0b011): comb
+= op
.eq(Op
.OP_BAD
)
928 with m
.Case(0b111): comb
+= op
.eq(Op
.OP_BAD
)
929 comb
+= req_op
.eq(op
)
930 comb
+= req_go
.eq(go
)
932 # Version of the row number that is valid one cycle earlier
933 # in the cases where we need to read the cache data BRAM.
934 # If we're stalling then we need to keep reading the last
936 with m
.If(~r0_stall
):
937 with m
.If(m_in
.valid
):
938 comb
+= early_req_row
.eq(get_row(m_in
.addr
))
940 comb
+= early_req_row
.eq(get_row(d_in
.addr
))
942 comb
+= early_req_row
.eq(req_row
)
944 def reservation_comb(self
, m
, cancel_store
, set_rsrv
, clear_rsrv
,
945 r0_valid
, r0
, reservation
):
946 """Handle load-with-reservation and store-conditional instructions
950 with m
.If(r0_valid
& r0
.req
.reserve
):
951 # XXX generate alignment interrupt if address
952 # is not aligned XXX or if r0.req.nc = '1'
953 with m
.If(r0
.req
.load
):
954 comb
+= set_rsrv
.eq(1) # load with reservation
956 comb
+= clear_rsrv
.eq(1) # store conditional
957 with m
.If((~reservation
.valid
) |
958 (r0
.req
.addr
[LINE_OFF_BITS
:64] != reservation
.addr
)):
959 comb
+= cancel_store
.eq(1)
961 def reservation_reg(self
, m
, r0_valid
, access_ok
, set_rsrv
, clear_rsrv
,
967 with m
.If(r0_valid
& access_ok
):
968 with m
.If(clear_rsrv
):
969 sync
+= reservation
.valid
.eq(0)
970 with m
.Elif(set_rsrv
):
971 sync
+= reservation
.valid
.eq(1)
972 sync
+= reservation
.addr
.eq(r0
.req
.addr
[LINE_OFF_BITS
:64])
974 def writeback_control(self
, m
, r1
, cache_out_row
):
975 """Return data for loads & completion control logic
979 d_out
, m_out
= self
.d_out
, self
.m_out
981 data_out
= Signal(64)
982 data_fwd
= Signal(64)
984 # Use the bypass if are reading the row that was
985 # written 1 or 2 cycles ago, including for the
986 # slow_valid = 1 case (i.e. completing a load
987 # miss or a non-cacheable load).
988 with m
.If(r1
.use_forward1
):
989 comb
+= data_fwd
.eq(r1
.forward_data1
)
991 comb
+= data_fwd
.eq(r1
.forward_data2
)
993 comb
+= data_out
.eq(cache_out_row
)
996 with m
.If(r1
.forward_sel
[i
]):
997 dsel
= data_fwd
.word_select(i
, 8)
998 comb
+= data_out
.word_select(i
, 8).eq(dsel
)
1000 comb
+= d_out
.valid
.eq(r1
.ls_valid
)
1001 comb
+= d_out
.data
.eq(data_out
)
1002 comb
+= d_out
.store_done
.eq(~r1
.stcx_fail
)
1003 comb
+= d_out
.error
.eq(r1
.ls_error
)
1004 comb
+= d_out
.cache_paradox
.eq(r1
.cache_paradox
)
1007 comb
+= m_out
.done
.eq(r1
.mmu_done
)
1008 comb
+= m_out
.err
.eq(r1
.mmu_error
)
1009 comb
+= m_out
.data
.eq(data_out
)
1011 # We have a valid load or store hit or we just completed
1012 # a slow op such as a load miss, a NC load or a store
1014 # Note: the load hit is delayed by one cycle. However it
1015 # can still not collide with r.slow_valid (well unless I
1016 # miscalculated) because slow_valid can only be set on a
1017 # subsequent request and not on its first cycle (the state
1018 # machine must have advanced), which makes slow_valid
1019 # at least 2 cycles from the previous hit_load_valid.
1021 # Sanity: Only one of these must be set in any given cycle
1023 if False: # TODO: need Display to get this to work
1024 assert (r1
.slow_valid
& r1
.stcx_fail
) != 1, \
1025 "unexpected slow_valid collision with stcx_fail"
1027 assert ((r1
.slow_valid | r1
.stcx_fail
) | r1
.hit_load_valid
) != 1, \
1028 "unexpected hit_load_delayed collision with slow_valid"
1030 with m
.If(~r1
.mmu_req
):
1031 # Request came from loadstore1...
1032 # Load hit case is the standard path
1033 with m
.If(r1
.hit_load_valid
):
1034 sync
+= Display("completing load hit data=%x", data_out
)
1036 # error cases complete without stalling
1037 with m
.If(r1
.ls_error
):
1038 sync
+= Display("completing ld/st with error")
1040 # Slow ops (load miss, NC, stores)
1041 with m
.If(r1
.slow_valid
):
1042 sync
+= Display("completing store or load miss adr=%x data=%x",
1043 r1
.req
.real_addr
, data_out
)
1046 # Request came from MMU
1047 with m
.If(r1
.hit_load_valid
):
1048 sync
+= Display("completing load hit to MMU, data=%x",
1050 # error cases complete without stalling
1051 with m
.If(r1
.mmu_error
):
1052 sync
+= Display("combpleting MMU ld with error")
1054 # Slow ops (i.e. load miss)
1055 with m
.If(r1
.slow_valid
):
1056 sync
+= Display("completing MMU load miss, data=%x",
1059 def rams(self
, m
, r1
, early_req_row
, cache_out_row
, replace_way
):
1061 Generate a cache RAM for each way. This handles the normal
1062 reads, writes from reloads and the special store-hit update
1065 Note: the BRAMs have an extra read buffer, meaning the output
1066 is pipelined an extra cycle. This differs from the
1067 icache. The writeback logic needs to take that into
1068 account by using 1-cycle delayed signals for load hits.
1073 for i
in range(NUM_WAYS
):
1074 do_read
= Signal(name
="do_rd%d" % i
)
1075 rd_addr
= Signal(ROW_BITS
)
1076 do_write
= Signal(name
="do_wr%d" % i
)
1077 wr_addr
= Signal(ROW_BITS
)
1078 wr_data
= Signal(WB_DATA_BITS
)
1079 wr_sel
= Signal(ROW_SIZE
)
1080 wr_sel_m
= Signal(ROW_SIZE
)
1081 _d_out
= Signal(WB_DATA_BITS
, name
="dout_%d" % i
) # cache_row_t
1083 way
= CacheRam(ROW_BITS
, WB_DATA_BITS
, ADD_BUF
=True)
1084 setattr(m
.submodules
, "cacheram_%d" % i
, way
)
1086 comb
+= way
.rd_en
.eq(do_read
)
1087 comb
+= way
.rd_addr
.eq(rd_addr
)
1088 comb
+= _d_out
.eq(way
.rd_data_o
)
1089 comb
+= way
.wr_sel
.eq(wr_sel_m
)
1090 comb
+= way
.wr_addr
.eq(wr_addr
)
1091 comb
+= way
.wr_data
.eq(wr_data
)
1094 comb
+= do_read
.eq(1)
1095 comb
+= rd_addr
.eq(early_req_row
)
1096 with m
.If(r1
.hit_way
== i
):
1097 comb
+= cache_out_row
.eq(_d_out
)
1101 # Defaults to wishbone read responses (cache refill)
1103 # For timing, the mux on wr_data/sel/addr is not
1104 # dependent on anything other than the current state.
1106 with m
.If(r1
.write_bram
):
1107 # Write store data to BRAM. This happens one
1108 # cycle after the store is in r0.
1109 comb
+= wr_data
.eq(r1
.req
.data
)
1110 comb
+= wr_sel
.eq(r1
.req
.byte_sel
)
1111 comb
+= wr_addr
.eq(get_row(r1
.req
.real_addr
))
1113 with m
.If(i
== r1
.req
.hit_way
):
1114 comb
+= do_write
.eq(1)
1116 # Otherwise, we might be doing a reload or a DCBZ
1118 comb
+= wr_data
.eq(0)
1120 comb
+= wr_data
.eq(wb_in
.dat
)
1121 comb
+= wr_addr
.eq(r1
.store_row
)
1122 comb
+= wr_sel
.eq(~
0) # all 1s
1124 with m
.If((r1
.state
== State
.RELOAD_WAIT_ACK
)
1125 & wb_in
.ack
& (replace_way
== i
)):
1126 comb
+= do_write
.eq(1)
1128 # Mask write selects with do_write since BRAM
1129 # doesn't have a global write-enable
1130 with m
.If(do_write
):
1131 comb
+= wr_sel_m
.eq(wr_sel
)
1133 # Cache hit synchronous machine for the easy case.
1134 # This handles load hits.
1135 # It also handles error cases (TLB miss, cache paradox)
1136 def dcache_fast_hit(self
, m
, req_op
, r0_valid
, r0
, r1
,
1137 req_hit_way
, req_index
, req_tag
, access_ok
,
1138 tlb_hit
, tlb_hit_way
, tlb_req_index
):
1143 with m
.If(req_op
!= Op
.OP_NONE
):
1144 sync
+= Display("op:%d addr:%x nc: %d idx: %x tag: %x way: %x",
1145 req_op
, r0
.req
.addr
, r0
.req
.nc
,
1146 req_index
, req_tag
, req_hit_way
)
1148 with m
.If(r0_valid
):
1149 sync
+= r1
.mmu_req
.eq(r0
.mmu_req
)
1151 # Fast path for load/store hits.
1152 # Set signals for the writeback controls.
1153 sync
+= r1
.hit_way
.eq(req_hit_way
)
1154 sync
+= r1
.hit_index
.eq(req_index
)
1156 with m
.If(req_op
== Op
.OP_LOAD_HIT
):
1157 sync
+= r1
.hit_load_valid
.eq(1)
1159 sync
+= r1
.hit_load_valid
.eq(0)
1161 with m
.If((req_op
== Op
.OP_LOAD_HIT
) |
(req_op
== Op
.OP_STORE_HIT
)):
1162 sync
+= r1
.cache_hit
.eq(1)
1164 sync
+= r1
.cache_hit
.eq(0)
1166 with m
.If(req_op
== Op
.OP_BAD
):
1167 # Display(f"Signalling ld/st error valid_ra={valid_ra}"
1168 # f"rc_ok={rc_ok} perm_ok={perm_ok}"
1169 sync
+= r1
.ls_error
.eq(~r0
.mmu_req
)
1170 sync
+= r1
.mmu_error
.eq(r0
.mmu_req
)
1171 sync
+= r1
.cache_paradox
.eq(access_ok
)
1174 sync
+= r1
.ls_error
.eq(0)
1175 sync
+= r1
.mmu_error
.eq(0)
1176 sync
+= r1
.cache_paradox
.eq(0)
1178 with m
.If(req_op
== Op
.OP_STCX_FAIL
):
1179 sync
+= r1
.stcx_fail
.eq(1)
1181 sync
+= r1
.stcx_fail
.eq(0)
1183 # Record TLB hit information for updating TLB PLRU
1184 sync
+= r1
.tlb_hit
.eq(tlb_hit
)
1185 sync
+= r1
.tlb_hit_way
.eq(tlb_hit_way
)
1186 sync
+= r1
.tlb_hit_index
.eq(tlb_req_index
)
1188 # Memory accesses are handled by this state machine:
1190 # * Cache load miss/reload (in conjunction with "rams")
1191 # * Load hits for non-cachable forms
1192 # * Stores (the collision case is handled in "rams")
1194 # All wishbone requests generation is done here.
1195 # This machine operates at stage 1.
1196 def dcache_slow(self
, m
, r1
, use_forward1_next
, use_forward2_next
,
1197 cache_valids
, r0
, replace_way
,
1198 req_hit_way
, req_same_tag
,
1199 r0_valid
, req_op
, cache_tags
, req_go
, ra
):
1205 req
= MemAccessRequest("mreq_ds")
1207 req_row
= Signal(ROW_BITS
)
1208 req_idx
= Signal(INDEX_BITS
)
1209 req_tag
= Signal(TAG_BITS
)
1210 comb
+= req_idx
.eq(get_index(req
.real_addr
))
1211 comb
+= req_row
.eq(get_row(req
.real_addr
))
1212 comb
+= req_tag
.eq(get_tag(req
.real_addr
))
1214 sync
+= r1
.use_forward1
.eq(use_forward1_next
)
1215 sync
+= r1
.forward_sel
.eq(0)
1217 with m
.If(use_forward1_next
):
1218 sync
+= r1
.forward_sel
.eq(r1
.req
.byte_sel
)
1219 with m
.Elif(use_forward2_next
):
1220 sync
+= r1
.forward_sel
.eq(r1
.forward_sel1
)
1222 sync
+= r1
.forward_data2
.eq(r1
.forward_data1
)
1223 with m
.If(r1
.write_bram
):
1224 sync
+= r1
.forward_data1
.eq(r1
.req
.data
)
1225 sync
+= r1
.forward_sel1
.eq(r1
.req
.byte_sel
)
1226 sync
+= r1
.forward_way1
.eq(r1
.req
.hit_way
)
1227 sync
+= r1
.forward_row1
.eq(get_row(r1
.req
.real_addr
))
1228 sync
+= r1
.forward_valid1
.eq(1)
1231 sync
+= r1
.forward_data1
.eq(0)
1233 sync
+= r1
.forward_data1
.eq(wb_in
.dat
)
1234 sync
+= r1
.forward_sel1
.eq(~
0) # all 1s
1235 sync
+= r1
.forward_way1
.eq(replace_way
)
1236 sync
+= r1
.forward_row1
.eq(r1
.store_row
)
1237 sync
+= r1
.forward_valid1
.eq(0)
1239 # One cycle pulses reset
1240 sync
+= r1
.slow_valid
.eq(0)
1241 sync
+= r1
.write_bram
.eq(0)
1242 sync
+= r1
.inc_acks
.eq(0)
1243 sync
+= r1
.dec_acks
.eq(0)
1245 sync
+= r1
.ls_valid
.eq(0)
1246 # complete tlbies and TLB loads in the third cycle
1247 sync
+= r1
.mmu_done
.eq(r0_valid
& (r0
.tlbie | r0
.tlbld
))
1249 with m
.If((req_op
== Op
.OP_LOAD_HIT
) |
(req_op
== Op
.OP_STCX_FAIL
)):
1250 with m
.If(~r0
.mmu_req
):
1251 sync
+= r1
.ls_valid
.eq(1)
1253 sync
+= r1
.mmu_done
.eq(1)
1255 with m
.If(r1
.write_tag
):
1256 # Store new tag in selected way
1257 for i
in range(NUM_WAYS
):
1258 with m
.If(i
== replace_way
):
1259 ct
= Signal(TAG_RAM_WIDTH
)
1260 comb
+= ct
.eq(cache_tags
[r1
.store_index
])
1263 cache_tags(r1.store_index)((i + 1) * TAG_WIDTH - 1 downto i * TAG_WIDTH) <=
1264 (TAG_WIDTH - 1 downto TAG_BITS => '0') & r1.reload_tag;
1266 comb
+= ct
.word_select(i
, TAG_WIDTH
).eq(r1
.reload_tag
)
1267 sync
+= cache_tags
[r1
.store_index
].eq(ct
)
1268 sync
+= r1
.store_way
.eq(replace_way
)
1269 sync
+= r1
.write_tag
.eq(0)
1271 # Take request from r1.req if there is one there,
1272 # else from req_op, ra, etc.
1274 comb
+= req
.eq(r1
.req
)
1276 comb
+= req
.op
.eq(req_op
)
1277 comb
+= req
.valid
.eq(req_go
)
1278 comb
+= req
.mmu_req
.eq(r0
.mmu_req
)
1279 comb
+= req
.dcbz
.eq(r0
.req
.dcbz
)
1280 comb
+= req
.real_addr
.eq(ra
)
1282 with m
.If(~r0
.req
.dcbz
):
1283 comb
+= req
.data
.eq(r0
.req
.data
)
1285 comb
+= req
.data
.eq(0)
1287 # Select all bytes for dcbz
1288 # and for cacheable loads
1289 with m
.If(r0
.req
.dcbz |
(r0
.req
.load
& ~r0
.req
.nc
)):
1290 comb
+= req
.byte_sel
.eq(~
0) # all 1s
1292 comb
+= req
.byte_sel
.eq(r0
.req
.byte_sel
)
1293 comb
+= req
.hit_way
.eq(req_hit_way
)
1294 comb
+= req
.same_tag
.eq(req_same_tag
)
1296 # Store the incoming request from r0,
1297 # if it is a slow request
1298 # Note that r1.full = 1 implies req_op = OP_NONE
1299 with m
.If((req_op
== Op
.OP_LOAD_MISS
)
1300 |
(req_op
== Op
.OP_LOAD_NC
)
1301 |
(req_op
== Op
.OP_STORE_MISS
)
1302 |
(req_op
== Op
.OP_STORE_HIT
)):
1303 sync
+= r1
.req
.eq(req
)
1304 sync
+= r1
.full
.eq(1)
1306 # Main state machine
1307 with m
.Switch(r1
.state
):
1309 with m
.Case(State
.IDLE
):
1310 sync
+= r1
.real_adr
.eq(req
.real_addr
)
1311 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1312 sync
+= r1
.wb
.dat
.eq(req
.data
)
1313 sync
+= r1
.dcbz
.eq(req
.dcbz
)
1315 # Keep track of our index and way
1316 # for subsequent stores.
1317 sync
+= r1
.store_index
.eq(req_idx
)
1318 sync
+= r1
.store_row
.eq(req_row
)
1319 sync
+= r1
.end_row_ix
.eq(get_row_of_line(req_row
)-1)
1320 sync
+= r1
.reload_tag
.eq(req_tag
)
1321 sync
+= r1
.req
.same_tag
.eq(1)
1323 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1324 sync
+= r1
.store_way
.eq(req
.hit_way
)
1326 # Reset per-row valid bits,
1327 # ready for handling OP_LOAD_MISS
1328 for i
in range(ROW_PER_LINE
):
1329 sync
+= r1
.rows_valid
[i
].eq(0)
1331 with m
.If(req_op
!= Op
.OP_NONE
):
1332 sync
+= Display("cache op %d", req
.op
)
1334 with m
.Switch(req
.op
):
1335 with m
.Case(Op
.OP_LOAD_HIT
):
1336 # stay in IDLE state
1339 with m
.Case(Op
.OP_LOAD_MISS
):
1340 sync
+= Display("cache miss real addr: %x " \
1342 req
.real_addr
, req_row
, req_tag
)
1344 # Start the wishbone cycle
1345 sync
+= r1
.wb
.we
.eq(0)
1346 sync
+= r1
.wb
.cyc
.eq(1)
1347 sync
+= r1
.wb
.stb
.eq(1)
1349 # Track that we had one request sent
1350 sync
+= r1
.state
.eq(State
.RELOAD_WAIT_ACK
)
1351 sync
+= r1
.write_tag
.eq(1)
1353 with m
.Case(Op
.OP_LOAD_NC
):
1354 sync
+= r1
.wb
.cyc
.eq(1)
1355 sync
+= r1
.wb
.stb
.eq(1)
1356 sync
+= r1
.wb
.we
.eq(0)
1357 sync
+= r1
.state
.eq(State
.NC_LOAD_WAIT_ACK
)
1359 with m
.Case(Op
.OP_STORE_HIT
, Op
.OP_STORE_MISS
):
1360 with m
.If(~req
.dcbz
):
1361 sync
+= r1
.state
.eq(State
.STORE_WAIT_ACK
)
1362 sync
+= r1
.acks_pending
.eq(1)
1363 sync
+= r1
.full
.eq(0)
1364 sync
+= r1
.slow_valid
.eq(1)
1366 with m
.If(~req
.mmu_req
):
1367 sync
+= r1
.ls_valid
.eq(1)
1369 sync
+= r1
.mmu_done
.eq(1)
1371 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1372 sync
+= r1
.write_bram
.eq(1)
1374 # dcbz is handled much like a load miss except
1375 # that we are writing to memory instead of reading
1376 sync
+= r1
.state
.eq(State
.RELOAD_WAIT_ACK
)
1378 with m
.If(req
.op
== Op
.OP_STORE_MISS
):
1379 sync
+= r1
.write_tag
.eq(1)
1381 sync
+= r1
.wb
.we
.eq(1)
1382 sync
+= r1
.wb
.cyc
.eq(1)
1383 sync
+= r1
.wb
.stb
.eq(1)
1385 # OP_NONE and OP_BAD do nothing
1386 # OP_BAD & OP_STCX_FAIL were
1387 # handled above already
1388 with m
.Case(Op
.OP_NONE
):
1390 with m
.Case(Op
.OP_BAD
):
1392 with m
.Case(Op
.OP_STCX_FAIL
):
1395 with m
.Case(State
.RELOAD_WAIT_ACK
):
1396 ld_stbs_done
= Signal()
1397 # Requests are all sent if stb is 0
1398 comb
+= ld_stbs_done
.eq(~r1
.wb
.stb
)
1400 with m
.If((~wb_in
.stall
) & r1
.wb
.stb
):
1401 # That was the last word? We are done sending.
1402 # Clear stb and set ld_stbs_done so we can handle an
1403 # eventual last ack on the same cycle.
1404 with m
.If(is_last_row_addr(r1
.real_adr
, r1
.end_row_ix
)):
1405 sync
+= r1
.wb
.stb
.eq(0)
1406 comb
+= ld_stbs_done
.eq(1)
1408 # Calculate the next row address in the current cache line
1409 row
= Signal(LINE_OFF_BITS
-ROW_OFF_BITS
)
1410 comb
+= row
.eq(r1
.real_adr
[ROW_OFF_BITS
:])
1411 sync
+= r1
.real_adr
[ROW_OFF_BITS
:LINE_OFF_BITS
].eq(row
+1)
1413 # Incoming acks processing
1414 sync
+= r1
.forward_valid1
.eq(wb_in
.ack
)
1415 with m
.If(wb_in
.ack
):
1416 srow
= Signal(ROW_LINE_BITS
)
1417 comb
+= srow
.eq(r1
.store_row
)
1418 sync
+= r1
.rows_valid
[srow
].eq(1)
1420 # If this is the data we were looking for,
1421 # we can complete the request next cycle.
1422 # Compare the whole address in case the
1423 # request in r1.req is not the one that
1424 # started this refill.
1425 with m
.If(r1
.full
& r1
.req
.same_tag
&
1426 ((r1
.dcbz
& r1
.req
.dcbz
) |
1427 ((~r1
.dcbz
) & (r1
.req
.op
== Op
.OP_LOAD_MISS
))) &
1428 (r1
.store_row
== get_row(r1
.req
.real_addr
))):
1429 sync
+= r1
.full
.eq(0)
1430 sync
+= r1
.slow_valid
.eq(1)
1431 with m
.If(~r1
.mmu_req
):
1432 sync
+= r1
.ls_valid
.eq(1)
1434 sync
+= r1
.mmu_done
.eq(1)
1435 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1436 sync
+= r1
.use_forward1
.eq(1)
1438 # Check for completion
1439 with m
.If(ld_stbs_done
& is_last_row(r1
.store_row
,
1441 # Complete wishbone cycle
1442 sync
+= r1
.wb
.cyc
.eq(0)
1444 # Cache line is now valid
1445 cv
= Signal(INDEX_BITS
)
1446 comb
+= cv
.eq(cache_valids
[r1
.store_index
])
1447 comb
+= cv
.bit_select(r1
.store_way
, 1).eq(1)
1448 sync
+= cache_valids
[r1
.store_index
].eq(cv
)
1450 sync
+= r1
.state
.eq(State
.IDLE
)
1452 # Increment store row counter
1453 sync
+= r1
.store_row
.eq(next_row(r1
.store_row
))
1455 with m
.Case(State
.STORE_WAIT_ACK
):
1456 st_stbs_done
= Signal()
1458 adjust_acks
= Signal(3)
1460 comb
+= st_stbs_done
.eq(~r1
.wb
.stb
)
1461 comb
+= acks
.eq(r1
.acks_pending
)
1463 with m
.If(r1
.inc_acks
!= r1
.dec_acks
):
1464 with m
.If(r1
.inc_acks
):
1465 comb
+= adjust_acks
.eq(acks
+ 1)
1467 comb
+= adjust_acks
.eq(acks
- 1)
1469 comb
+= adjust_acks
.eq(acks
)
1471 sync
+= r1
.acks_pending
.eq(adjust_acks
)
1473 # Clear stb when slave accepted request
1474 with m
.If(~wb_in
.stall
):
1475 # See if there is another store waiting
1476 # to be done which is in the same real page.
1477 with m
.If(req
.valid
):
1478 ra
= req
.real_addr
[0:SET_SIZE_BITS
]
1479 sync
+= r1
.real_adr
[0:SET_SIZE_BITS
].eq(ra
)
1480 sync
+= r1
.wb
.dat
.eq(req
.data
)
1481 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1483 with m
.If((adjust_acks
< 7) & req
.same_tag
&
1484 ((req
.op
== Op
.OP_STORE_MISS
)
1485 |
(req
.op
== Op
.OP_STORE_HIT
))):
1486 sync
+= r1
.wb
.stb
.eq(1)
1487 comb
+= st_stbs_done
.eq(0)
1489 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1490 sync
+= r1
.write_bram
.eq(1)
1491 sync
+= r1
.full
.eq(0)
1492 sync
+= r1
.slow_valid
.eq(1)
1494 # Store requests never come from the MMU
1495 sync
+= r1
.ls_valid
.eq(1)
1496 comb
+= st_stbs_done
.eq(0)
1497 sync
+= r1
.inc_acks
.eq(1)
1499 sync
+= r1
.wb
.stb
.eq(0)
1500 comb
+= st_stbs_done
.eq(1)
1502 # Got ack ? See if complete.
1503 with m
.If(wb_in
.ack
):
1504 with m
.If(st_stbs_done
& (adjust_acks
== 1)):
1505 sync
+= r1
.state
.eq(State
.IDLE
)
1506 sync
+= r1
.wb
.cyc
.eq(0)
1507 sync
+= r1
.wb
.stb
.eq(0)
1508 sync
+= r1
.dec_acks
.eq(1)
1510 with m
.Case(State
.NC_LOAD_WAIT_ACK
):
1511 # Clear stb when slave accepted request
1512 with m
.If(~wb_in
.stall
):
1513 sync
+= r1
.wb
.stb
.eq(0)
1515 # Got ack ? complete.
1516 with m
.If(wb_in
.ack
):
1517 sync
+= r1
.state
.eq(State
.IDLE
)
1518 sync
+= r1
.full
.eq(0)
1519 sync
+= r1
.slow_valid
.eq(1)
1521 with m
.If(~r1
.mmu_req
):
1522 sync
+= r1
.ls_valid
.eq(1)
1524 sync
+= r1
.mmu_done
.eq(1)
1526 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1527 sync
+= r1
.use_forward1
.eq(1)
1528 sync
+= r1
.wb
.cyc
.eq(0)
1529 sync
+= r1
.wb
.stb
.eq(0)
1531 def dcache_log(self
, m
, r1
, valid_ra
, tlb_hit_way
, stall_out
):
1534 d_out
, wb_in
, log_out
= self
.d_out
, self
.wb_in
, self
.log_out
1536 sync
+= log_out
.eq(Cat(r1
.state
[:3], valid_ra
, tlb_hit_way
[:3],
1537 stall_out
, req_op
[:3], d_out
.valid
, d_out
.error
,
1538 r1
.wb
.cyc
, r1
.wb
.stb
, wb_in
.ack
, wb_in
.stall
,
1541 def elaborate(self
, platform
):
1546 # Storage. Hopefully "cache_rows" is a BRAM, the rest is LUTs
1547 cache_tags
= CacheTagArray()
1548 cache_tag_set
= Signal(TAG_RAM_WIDTH
)
1549 cache_valids
= CacheValidBitsArray()
1551 # TODO attribute ram_style : string;
1552 # TODO attribute ram_style of cache_tags : signal is "distributed";
1554 """note: these are passed to nmigen.hdl.Memory as "attributes".
1555 don't know how, just that they are.
1557 dtlb_valid_bits
= TLBValidBitsArray()
1558 dtlb_tags
= TLBTagsArray()
1559 dtlb_ptes
= TLBPtesArray()
1560 # TODO attribute ram_style of
1561 # dtlb_tags : signal is "distributed";
1562 # TODO attribute ram_style of
1563 # dtlb_ptes : signal is "distributed";
1565 r0
= RegStage0("r0")
1568 r1
= RegStage1("r1")
1570 reservation
= Reservation()
1572 # Async signals on incoming request
1573 req_index
= Signal(INDEX_BITS
)
1574 req_row
= Signal(ROW_BITS
)
1575 req_hit_way
= Signal(WAY_BITS
)
1576 req_tag
= Signal(TAG_BITS
)
1578 req_data
= Signal(64)
1579 req_same_tag
= Signal()
1582 early_req_row
= Signal(ROW_BITS
)
1584 cancel_store
= Signal()
1586 clear_rsrv
= Signal()
1591 use_forward1_next
= Signal()
1592 use_forward2_next
= Signal()
1594 cache_out_row
= Signal(WB_DATA_BITS
)
1596 plru_victim
= PLRUOut()
1597 replace_way
= Signal(WAY_BITS
)
1599 # Wishbone read/write/cache write formatting signals
1603 tlb_tag_way
= Signal(TLB_TAG_WAY_BITS
)
1604 tlb_pte_way
= Signal(TLB_PTE_WAY_BITS
)
1605 tlb_valid_way
= Signal(TLB_NUM_WAYS
)
1606 tlb_req_index
= Signal(TLB_SET_BITS
)
1608 tlb_hit_way
= Signal(TLB_WAY_BITS
)
1609 pte
= Signal(TLB_PTE_BITS
)
1610 ra
= Signal(REAL_ADDR_BITS
)
1612 perm_attr
= PermAttr("dc_perms")
1615 access_ok
= Signal()
1617 tlb_plru_victim
= TLBPLRUOut()
1619 # we don't yet handle collisions between loadstore1 requests
1621 comb
+= self
.m_out
.stall
.eq(0)
1623 # Hold off the request in r0 when r1 has an uncompleted request
1624 comb
+= r0_stall
.eq(r0_full
& r1
.full
)
1625 comb
+= r0_valid
.eq(r0_full
& ~r1
.full
)
1626 comb
+= self
.stall_out
.eq(r0_stall
)
1628 # Wire up wishbone request latch out of stage 1
1629 comb
+= r1
.wb
.adr
.eq(r1
.real_adr
)
1630 comb
+= self
.wb_out
.eq(r1
.wb
)
1631 comb
+= self
.wb_out
.adr
.eq(r1
.wb
.adr
[3:]) # truncate LSBs
1633 # deal with litex not doing wishbone pipeline mode
1634 comb
+= self
.wb_in
.stall
.eq(self
.wb_out
.cyc
& ~self
.wb_in
.ack
)
1636 # call sub-functions putting everything together, using shared
1637 # signals established above
1638 self
.stage_0(m
, r0
, r1
, r0_full
)
1639 self
.tlb_read(m
, r0_stall
, tlb_valid_way
,
1640 tlb_tag_way
, tlb_pte_way
, dtlb_valid_bits
,
1641 dtlb_tags
, dtlb_ptes
)
1642 self
.tlb_search(m
, tlb_req_index
, r0
, r0_valid
,
1643 tlb_valid_way
, tlb_tag_way
, tlb_hit_way
,
1644 tlb_pte_way
, pte
, tlb_hit
, valid_ra
, perm_attr
, ra
)
1645 self
.tlb_update(m
, r0_valid
, r0
, dtlb_valid_bits
, tlb_req_index
,
1646 tlb_hit_way
, tlb_hit
, tlb_plru_victim
, tlb_tag_way
,
1647 dtlb_tags
, tlb_pte_way
, dtlb_ptes
)
1648 self
.maybe_plrus(m
, r1
, plru_victim
)
1649 self
.maybe_tlb_plrus(m
, r1
, tlb_plru_victim
)
1650 self
.cache_tag_read(m
, r0_stall
, req_index
, cache_tag_set
, cache_tags
)
1651 self
.dcache_request(m
, r0
, ra
, req_index
, req_row
, req_tag
,
1652 r0_valid
, r1
, cache_valids
, replace_way
,
1653 use_forward1_next
, use_forward2_next
,
1654 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
1655 valid_ra
, perm_ok
, access_ok
, req_op
, req_go
,
1657 tlb_hit
, tlb_hit_way
, tlb_valid_way
, cache_tag_set
,
1658 cancel_store
, req_same_tag
, r0_stall
, early_req_row
)
1659 self
.reservation_comb(m
, cancel_store
, set_rsrv
, clear_rsrv
,
1660 r0_valid
, r0
, reservation
)
1661 self
.reservation_reg(m
, r0_valid
, access_ok
, set_rsrv
, clear_rsrv
,
1663 self
.writeback_control(m
, r1
, cache_out_row
)
1664 self
.rams(m
, r1
, early_req_row
, cache_out_row
, replace_way
)
1665 self
.dcache_fast_hit(m
, req_op
, r0_valid
, r0
, r1
,
1666 req_hit_way
, req_index
, req_tag
, access_ok
,
1667 tlb_hit
, tlb_hit_way
, tlb_req_index
)
1668 self
.dcache_slow(m
, r1
, use_forward1_next
, use_forward2_next
,
1669 cache_valids
, r0
, replace_way
,
1670 req_hit_way
, req_same_tag
,
1671 r0_valid
, req_op
, cache_tags
, req_go
, ra
)
1672 #self.dcache_log(m, r1, valid_ra, tlb_hit_way, stall_out)
1676 def dcache_load(dut
, addr
, nc
=0):
1677 yield dut
.d_in
.load
.eq(1)
1678 yield dut
.d_in
.nc
.eq(nc
)
1679 yield dut
.d_in
.addr
.eq(addr
)
1680 yield dut
.d_in
.byte_sel
.eq(~
0)
1681 yield dut
.d_in
.valid
.eq(1)
1683 yield dut
.d_in
.valid
.eq(0)
1684 yield dut
.d_in
.byte_sel
.eq(0)
1685 while not (yield dut
.d_out
.valid
):
1687 data
= yield dut
.d_out
.data
1691 def dcache_store(dut
, addr
, data
, nc
=0):
1692 yield dut
.d_in
.load
.eq(0)
1693 yield dut
.d_in
.nc
.eq(nc
)
1694 yield dut
.d_in
.data
.eq(data
)
1695 yield dut
.d_in
.byte_sel
.eq(~
0)
1696 yield dut
.d_in
.addr
.eq(addr
)
1697 yield dut
.d_in
.valid
.eq(1)
1699 yield dut
.d_in
.valid
.eq(0)
1700 yield dut
.d_in
.byte_sel
.eq(0)
1701 while not (yield dut
.d_out
.valid
):
1705 def dcache_random_sim(dut
, mem
):
1708 sim_mem
= deepcopy(mem
)
1709 memsize
= len(sim_mem
)
1710 print ("mem len", memsize
)
1713 yield dut
.d_in
.valid
.eq(0)
1714 yield dut
.d_in
.load
.eq(0)
1715 yield dut
.d_in
.priv_mode
.eq(1)
1716 yield dut
.d_in
.nc
.eq(0)
1717 yield dut
.d_in
.addr
.eq(0)
1718 yield dut
.d_in
.data
.eq(0)
1719 yield dut
.m_in
.valid
.eq(0)
1720 yield dut
.m_in
.addr
.eq(0)
1721 yield dut
.m_in
.pte
.eq(0)
1722 # wait 4 * clk_period
1730 #for i in range(1024):
1733 for i
in range(1024):
1734 addr
= randint(0, memsize
-1)
1735 data
= randint(0, (1<<64)-1)
1736 sim_mem
[addr
] = data
1740 print ("random testing %d 0x%x row %d data 0x%x" % (i
, addr
, row
, data
))
1742 yield from dcache_load(dut
, addr
)
1743 yield from dcache_store(dut
, addr
, data
)
1745 addr
= randint(0, memsize
-1)
1746 sim_data
= sim_mem
[addr
]
1750 print (" load 0x%x row %d expect data 0x%x" % (addr
, row
, sim_data
))
1751 data
= yield from dcache_load(dut
, addr
)
1752 assert data
== sim_data
, \
1753 "check addr 0x%x row %d data %x != %x" % (addr
, row
, data
, sim_data
)
1755 for addr
in range(memsize
):
1756 data
= yield from dcache_load(dut
, addr
*8)
1757 assert data
== sim_mem
[addr
], \
1758 "final check %x data %x != %x" % (addr
*8, data
, sim_mem
[addr
])
1761 def dcache_sim(dut
, mem
):
1763 yield dut
.d_in
.valid
.eq(0)
1764 yield dut
.d_in
.load
.eq(0)
1765 yield dut
.d_in
.priv_mode
.eq(1)
1766 yield dut
.d_in
.nc
.eq(0)
1767 yield dut
.d_in
.addr
.eq(0)
1768 yield dut
.d_in
.data
.eq(0)
1769 yield dut
.m_in
.valid
.eq(0)
1770 yield dut
.m_in
.addr
.eq(0)
1771 yield dut
.m_in
.pte
.eq(0)
1772 # wait 4 * clk_period
1778 # Cacheable read of address 4
1779 data
= yield from dcache_load(dut
, 0x58)
1780 addr
= yield dut
.d_in
.addr
1781 assert data
== 0x0000001700000016, \
1782 f
"data @%x=%x expected 0x0000001700000016" % (addr
, data
)
1784 # Cacheable read of address 20
1785 data
= yield from dcache_load(dut
, 0x20)
1786 addr
= yield dut
.d_in
.addr
1787 assert data
== 0x0000000900000008, \
1788 f
"data @%x=%x expected 0x0000000900000008" % (addr
, data
)
1790 # Cacheable read of address 30
1791 data
= yield from dcache_load(dut
, 0x530)
1792 addr
= yield dut
.d_in
.addr
1793 assert data
== 0x0000014D0000014C, \
1794 f
"data @%x=%x expected 0000014D0000014C" % (addr
, data
)
1796 # 2nd Cacheable read of address 30
1797 data
= yield from dcache_load(dut
, 0x530)
1798 addr
= yield dut
.d_in
.addr
1799 assert data
== 0x0000014D0000014C, \
1800 f
"data @%x=%x expected 0000014D0000014C" % (addr
, data
)
1802 # Non-cacheable read of address 100
1803 data
= yield from dcache_load(dut
, 0x100, nc
=1)
1804 addr
= yield dut
.d_in
.addr
1805 assert data
== 0x0000004100000040, \
1806 f
"data @%x=%x expected 0000004100000040" % (addr
, data
)
1808 # Store at address 530
1809 yield from dcache_store(dut
, 0x530, 0x121)
1811 # Store at address 30
1812 yield from dcache_store(dut
, 0x530, 0x12345678)
1814 # 3nd Cacheable read of address 530
1815 data
= yield from dcache_load(dut
, 0x530)
1816 addr
= yield dut
.d_in
.addr
1817 assert data
== 0x12345678, \
1818 f
"data @%x=%x expected 0x12345678" % (addr
, data
)
1820 # 4th Cacheable read of address 20
1821 data
= yield from dcache_load(dut
, 0x20)
1822 addr
= yield dut
.d_in
.addr
1823 assert data
== 0x0000000900000008, \
1824 f
"data @%x=%x expected 0x0000000900000008" % (addr
, data
)
1832 def test_dcache(mem
, test_fn
, test_name
):
1835 memory
= Memory(width
=64, depth
=len(mem
), init
=mem
, simulate
=True)
1836 sram
= SRAM(memory
=memory
, granularity
=8)
1839 m
.submodules
.dcache
= dut
1840 m
.submodules
.sram
= sram
1842 m
.d
.comb
+= sram
.bus
.cyc
.eq(dut
.wb_out
.cyc
)
1843 m
.d
.comb
+= sram
.bus
.stb
.eq(dut
.wb_out
.stb
)
1844 m
.d
.comb
+= sram
.bus
.we
.eq(dut
.wb_out
.we
)
1845 m
.d
.comb
+= sram
.bus
.sel
.eq(dut
.wb_out
.sel
)
1846 m
.d
.comb
+= sram
.bus
.adr
.eq(dut
.wb_out
.adr
)
1847 m
.d
.comb
+= sram
.bus
.dat_w
.eq(dut
.wb_out
.dat
)
1849 m
.d
.comb
+= dut
.wb_in
.ack
.eq(sram
.bus
.ack
)
1850 m
.d
.comb
+= dut
.wb_in
.dat
.eq(sram
.bus
.dat_r
)
1856 sim
.add_sync_process(wrap(test_fn(dut
, mem
)))
1857 with sim
.write_vcd('test_dcache%s.vcd' % test_name
):
1860 if __name__
== '__main__':
1863 vl
= rtlil
.convert(dut
, ports
=[])
1864 with
open("test_dcache.il", "w") as f
:
1868 for i
in range(1024):
1869 mem
.append((i
*2)|
((i
*2+1)<<32))
1871 test_dcache(mem
, dcache_sim
, "")
1875 for i
in range(memsize
):
1878 test_dcache(mem
, dcache_random_sim
, "random")