3 based on Anton Blanchard microwatt dcache.vhdl
7 from enum
import Enum
, unique
9 from nmigen
import Module
, Signal
, Elaboratable
,
11 from nmigen
.cli
import main
12 from nmigen
.iocontrol
import RecordObject
13 from nmigen
.util
import log2_int
15 from experiment
.mem_types
import LoadStore1ToDCacheType
,
16 DCacheToLoadStore1Type
,
20 from experiment
.wb_types
import WB_ADDR_BITS
, WB_DATA_BITS
, WB_SEL_BITS
,
21 WBAddrType
, WBDataType
, WBSelType
,
22 WbMasterOut
, WBSlaveOut
,
23 WBMasterOutVector
, WBSlaveOutVector
,
24 WBIOMasterOut
, WBIOSlaveOut
26 # TODO: make these parameters of DCache at some point
27 LINE_SIZE
= 64 # Line size in bytes
28 NUM_LINES
= 32 # Number of lines in a set
29 NUM_WAYS
= 4 # Number of ways
30 TLB_SET_SIZE
= 64 # L1 DTLB entries per set
31 TLB_NUM_WAYS
= 2 # L1 DTLB number of sets
32 TLB_LG_PGSZ
= 12 # L1 DTLB log_2(page_size)
33 LOG_LENGTH
= 0 # Non-zero to enable log data collection
35 # BRAM organisation: We never access more than
36 # -- wishbone_data_bits at a time so to save
37 # -- resources we make the array only that wide, and
38 # -- use consecutive indices for to make a cache "line"
40 # -- ROW_SIZE is the width in bytes of the BRAM
41 # -- (based on WB, so 64-bits)
42 ROW_SIZE
= WB_DATA_BITS
// 8;
44 # ROW_PER_LINE is the number of row (wishbone
45 # transactions) in a line
46 ROW_PER_LINE
= LINE_SIZE
// ROW_SIZE
48 # BRAM_ROWS is the number of rows in BRAM needed
49 # to represent the full dcache
50 BRAM_ROWS
= NUM_LINES
* ROW_PER_LINE
53 # Bit fields counts in the address
55 # REAL_ADDR_BITS is the number of real address
59 # ROW_BITS is the number of bits to select a row
60 ROW_BITS
= log2_int(BRAM_ROWS
)
62 # ROW_LINE_BITS is the number of bits to select
64 ROW_LINE_BITS
= log2_int(ROW_PER_LINE
)
66 # LINE_OFF_BITS is the number of bits for
67 # the offset in a cache line
68 LINE_OFF_BITS
= log2_int(LINE_SIZE
)
70 # ROW_OFF_BITS is the number of bits for
72 ROW_OFF_BITS
= log2_int(ROW_SIZE
)
74 # INDEX_BITS is the number if bits to
76 INDEX_BITS
= log2_int(NUM_LINES
)
78 # SET_SIZE_BITS is the log base 2 of the set size
79 SET_SIZE_BITS
= LINE_OFF_BITS
+ INDEX_BITS
81 # TAG_BITS is the number of bits of
82 # the tag part of the address
83 TAG_BITS
= REAL_ADDR_BITS
- SET_SIZE_BITS
85 # TAG_WIDTH is the width in bits of each way of the tag RAM
86 TAG_WIDTH
= TAG_BITS
+ 7 - ((TAG_BITS
+ 7) % 8)
88 # WAY_BITS is the number of bits to select a way
89 WAY_BITS
= log2_int(NUM_WAYS
)
91 # Example of layout for 32 lines of 64 bytes:
93 # .. tag |index| line |
95 # .. | |---| | ROW_LINE_BITS (3)
96 # .. | |--- - --| LINE_OFF_BITS (6)
97 # .. | |- --| ROW_OFF_BITS (3)
98 # .. |----- ---| | ROW_BITS (8)
99 # .. |-----| | INDEX_BITS (5)
100 # .. --------| | TAG_BITS (45)
102 TAG_RAM_WIDTH
= TAG_WIDTH
* NUM_WAYS
105 return Array(CacheTagSet() for x
in range(NUM_LINES
))
107 def CacheValidBitsArray():
108 return Array(CacheWayValidBits() for x
in range(NUM_LINES
))
110 def RowPerLineValidArray():
111 return Array(Signal() for x
in range(ROW_PER_LINE
))
114 TLB_SET_BITS
= log2_int(TLB_SET_SIZE
)
115 TLB_WAY_BITS
= log2_int(TLB_NUM_WAYS
)
116 TLB_EA_TAG_BITS
= 64 - (TLB_LG_PGSZ
+ TLB_SET_BITS
)
117 TLB_TAG_WAY_BITS
= TLB_NUM_WAYS
* TLB_EA_TAG_BITS
119 TLB_PTE_WAY_BITS
= TLB_NUM_WAYS
* TLB_PTE_BITS
;
121 assert (LINE_SIZE
% ROW_SIZE
) == 0, "LINE_SIZE not multiple of ROW_SIZE"
122 assert (LINE_SIZE
% 2) == 0, "LINE_SIZE not power of 2"
123 assert (NUM_LINES
% 2) == 0, "NUM_LINES not power of 2"
124 assert (ROW_PER_LINE
% 2) == 0, "ROW_PER_LINE not power of 2"
125 assert ROW_BITS
== (INDEX_BITS
+ ROW_LINE_BITS
), "geometry bits don't add up"
126 assert (LINE_OFF_BITS
= ROW_OFF_BITS
+ ROW_LINEBITS
), \
127 "geometry bits don't add up"
128 assert REAL_ADDR_BITS
== (TAG_BITS
+ INDEX_BITS
+ LINE_OFF_BITS
), \
129 "geometry bits don't add up"
130 assert REAL_ADDR_BITS
== (TAG_BITS
+ ROW_BITS
+ ROW_OFF_BITS
), \
131 "geometry bits don't add up"
132 assert 64 == wishbone_data_bits
, "Can't yet handle wb width that isn't 64-bits"
133 assert SET_SIZE_BITS
<= TLB_LG_PGSZ
, "Set indexed by virtual address"
136 def TLBValidBitsArray():
137 return Array(Signal(TLB_NUM_WAYS
) for x
in range(TLB_SET_SIZE
))
140 return Array(Signal(TLB_TAG_WAY_BITS
) for x
in range (TLB_SET_SIZE
))
143 return Array(Signal(TLB_PTE_WAY_BITS
) for x
in range(TLB_SET_SIZE
))
146 return Array(Signal(NUM_WAYS
) for x
in range(TLB_NUM_WAYS
))
148 # Cache RAM interface
150 return Array(Signal(WB_DATA_BITS
) for x
in range(NUM_WAYS
))
152 # PLRU output interface
154 return Array(Signal(WAY_BITS
) for x
in range(Index()))
156 # TLB PLRU output interface
158 return Array(Signal(TLB_WAY_BITS
) for x
in range(TLB_SET_SIZE
))
160 # Helper functions to decode incoming requests
162 # Return the cache line index (tag index) for an address
164 return addr
[LINE_OFF_BITS
:SET_SIZE_BITS
]
166 # Return the cache row index (data memory) for an address
168 return addr
[ROW_OFF_BITS
:SET_SIZE_BITS
]
170 # Return the index of a row within a line
171 def get_row_of_line(row
):
172 row_v
= Signal(ROW_BITS
)
174 return row_v
[0:ROW_LINE_BITS
]
176 # Returns whether this is the last row of a line
177 def is_last_row_addr(addr
, last
):
178 return addr
[ROW_OFF_BITS
:LINE_OFF_BITS
] == last
180 # Returns whether this is the last row of a line
181 def is_last_row(row
, last
):
182 return get_row_of_line(row
) == last
184 # Return the address of the next row in the current cache line
185 def next_row_addr(addr
):
186 row_idx
= Signal(ROW_LINE_BITS
)
187 result
= WBAddrType()
188 # Is there no simpler way in VHDL to
189 # generate that 3 bits adder ?
190 row_idx
= addr
[ROW_OFF_BITS
:LINE_OFF_BITS
]
191 row_idx
= Signal(row_idx
+ 1)
193 result
[ROW_OFF_BITS
:LINE_OFF_BITS
] = row_idx
196 # Return the next row in the current cache line. We use a
197 # dedicated function in order to limit the size of the
198 # generated adder to be only the bits within a cache line
199 # (3 bits with default settings)
201 row_v
= row
[0:ROW_LINE_BITS
] + 1
202 return Cat(row_v
[:ROW_LINE_BITS
], row
[ROW_LINE_BITS
:])
204 # Get the tag value from the address
206 return addr
[SET_SIZE_BITS
:REAL_ADDR_BITS
]
208 # Read a tag from a tag memory row
209 def read_tag(way
, tagset
):
210 return tagset
[way
*TAG_WIDTH
:way
* TAG_WIDTH
+ TAG_BITS
]
212 # Read a TLB tag from a TLB tag memory row
213 def read_tlb_tag(way
, tags
):
214 j
= way
* TLB_EA_TAG_BITS
215 return tags
[j
:j
+ TLB_EA_TAG_BITS
]
217 # Write a TLB tag to a TLB tag memory row
218 def write_tlb_tag(way
, tags
), tag
):
219 j
= way
* TLB_EA_TAG_BITS
220 tags
[j
:j
+ TLB_EA_TAG_BITS
] = tag
222 # Read a PTE from a TLB PTE memory row
223 def read_tlb_pte(way
, ptes
):
224 j
= way
* TLB_PTE_BITS
225 return ptes
[j
:j
+ TLB_PTE_BITS
]
227 def write_tlb_pte(way
, ptes
,newpte
):
228 j
= way
* TLB_PTE_BITS
229 return ptes
[j
:j
+ TLB_PTE_BITS
].eq(newpte
)
232 # Record for storing permission, attribute, etc. bits from a PTE
233 class PermAttr(RecordObject
):
236 self
.reference
= Signal()
237 self
.changed
= Signal()
238 self
.nocache
= Signal()
240 self
.rd_perm
= Signal()
241 self
.wr_perm
= Signal()
244 def extract_perm_attr(pte
):
246 pa
.reference
= pte
[8]
255 # Type of operation on a "valid" input
259 OP_BAD
= 1 # NC cache hit, TLB miss, prot/RC failure
260 OP_STCX_FAIL
= 2 # conditional store w/o reservation
261 OP_LOAD_HIT
= 3 # Cache hit on load
262 OP_LOAD_MISS
= 4 # Load missing cache
263 OP_LOAD_NC
= 5 # Non-cachable load
264 OP_STORE_HIT
= 6 # Store hitting cache
265 OP_STORE_MISS
= 7 # Store missing cache
268 # Cache state machine
271 IDLE
= 0 # Normal load hit processing
272 RELOAD_WAIT_ACK
= 1 # Cache reload wait ack
273 STORE_WAIT_ACK
= 2 # Store wait ack
274 NC_LOAD_WAIT_ACK
= 3 # Non-cachable load wait ack
279 # In order to make timing, we use the BRAMs with
280 # an output buffer, which means that the BRAM
281 # output is delayed by an extra cycle.
283 # Thus, the dcache has a 2-stage internal pipeline
284 # for cache hits with no stalls.
286 # All other operations are handled via stalling
287 # in the first stage.
289 # The second stage can thus complete a hit at the same
290 # time as the first stage emits a stall for a complex op.
292 # Stage 0 register, basically contains just the latched request
294 class RegStage0(RecordObject
):
297 self
.req
= LoadStore1ToDCacheType()
298 self
.tlbie
= Signal()
299 self
.doall
= Signal()
300 self
.tlbld
= Signal()
301 self
.mmu_req
= Signal() # indicates source of request
304 class MemAccessRequest(RecordObject
):
308 self
.valid
= Signal()
310 self
.real_addr
= Signal(REAL_ADDR_BITS
)
311 self
.data
= Signal(64)
312 self
.byte_sel
= Signal(8)
313 self
.hit_way
= Signal(WAY_BITS
)
314 self
.same_tag
= Signal()
315 self
.mmu_req
= Signal()
318 # First stage register, contains state for stage 1 of load hits
319 # and for the state machine used by all other operations
320 class RegStage1(RecordObject
):
323 # Info about the request
324 self
.full
= Signal() # have uncompleted request
325 self
.mmu_req
= Signal() # request is from MMU
326 self
.req
= MemAccessRequest()
329 self
.hit_way
= Signal(WAY_BITS
)
330 self
.hit_load_valid
= Signal()
331 self
.hit_index
= Signal(NUM_LINES
)
332 self
.cache_hit
= Signal()
335 self
.tlb_hit
= Signal()
336 self
.tlb_hit_way
= Signal(TLB_NUM_WAYS
)
337 self
.tlb_hit_index
= Signal(TLB_WAY_BITS
)
339 # 2-stage data buffer for data forwarded from writes to reads
340 self
.forward_data1
= Signal(64)
341 self
.forward_data2
= Signal(64)
342 self
.forward_sel1
= Signal(8)
343 self
.forward_valid1
= Signal()
344 self
.forward_way1
= Signal(WAY_BITS
)
345 self
.forward_row1
= Signal(ROW_BITS
)
346 self
.use_forward1
= Signal()
347 self
.forward_sel
= Signal(8)
349 # Cache miss state (reload state machine)
352 self
.write_bram
= Signal()
353 self
.write_tag
= Signal()
354 self
.slow_valid
= Signal()
355 self
.wb
= WishboneMasterOut()
356 self
.reload_tag
= Signal(TAG_BITS
)
357 self
.store_way
= Signal(WAY_BITS
)
358 self
.store_row
= Signal(ROW_BITS
)
359 self
.store_index
= Signal(INDEX_BITS
)
360 self
.end_row_ix
= Signal(log2_int(ROW_LINE_BITS
))
361 self
.rows_valid
= RowPerLineValidArray()
362 self
.acks_pending
= Signal(3)
363 self
.inc_acks
= Signal()
364 self
.dec_acks
= Signal()
366 # Signals to complete (possibly with error)
367 self
.ls_valid
= Signal()
368 self
.ls_error
= Signal()
369 self
.mmu_done
= Signal()
370 self
.mmu_error
= Signal()
371 self
.cache_paradox
= Signal()
373 # Signal to complete a failed stcx.
374 self
.stcx_fail
= Signal()
377 # Reservation information
378 class Reservation(RecordObject
):
381 self
.valid
= Signal()
382 self
.addr
= Signal(64-LINE_OFF_BITS
)
385 class DCache(Elaboratable
):
386 """Set associative dcache write-through
387 TODO (in no specific order):
388 * See list in icache.vhdl
389 * Complete load misses on the cycle when WB data comes instead of
390 at the end of line (this requires dealing with requests coming in
394 self
.d_in
= LoadStore1ToDCacheType()
395 self
.d_out
= DCacheToLoadStore1Type()
397 self
.m_in
= MMUToDCacheType()
398 self
.m_out
= DCacheToMMUType()
400 self
.stall_out
= Signal()
402 self
.wb_out
= WBMasterOut()
403 self
.wb_in
= WBSlaveOut()
405 self
.log_out
= Signal(20)
407 def stage_0(self
, m
, d_in
, m_in
):
408 """Latch the request in r0.req as long as we're not stalling
415 # TODO, this goes in unit tests and formal proofs
416 with m
.If(~
(d_in
.valid
& m_in
.valid
)):
417 #sync += Display("request collision loadstore vs MMU")
420 with m
.If(m_in
.valid
):
421 sync
+= r
.req
.valid
.eq(1)
422 sync
+= r
.req
.load
.eq(~
(m_in
.tlbie | m_in
.tlbld
))
423 sync
+= r
.req
.dcbz
.eq(0)
424 sync
+= r
.req
.nc
.eq(0)
425 sync
+= r
.req
.reserve
.eq(0)
426 sync
+= r
.req
.virt_mode
.eq(1)
427 sync
+= r
.req
.priv_mode
.eq(1)
428 sync
+= r
.req
.addr
.eq(m_in
.addr
)
429 sync
+= r
.req
.data
.eq(m_in
.pte
)
430 sync
+= r
.req
.byte_sel
.eq(~
0) # Const -1 sets all to 0b111....
431 sync
+= r
.tlbie
.eq(m_in
.tlbie
)
432 sync
+= r
.doall
.eq(m_in
.doall
)
433 sync
+= r
.tlbld
.eq(m_in
.tlbld
)
434 sync
+= r
.mmu_req
.eq(1)
436 sync
+= r
.req
.eq(d_in
)
437 sync
+= r
.req
.tlbie
.eq(0)
438 sync
+= r
.req
.doall
.eq(0)
439 sync
+= r
.req
.tlbd
.eq(0)
440 sync
+= r
.req
.mmu_req
.eq(0)
441 with m
.If(~
(r1
.full
& r0_full
)):
443 sync
+= r0_full
.eq(r
.req
.valid
)
445 def tlb_read(self
, m
, m_in
, d_in
, r0_stall
, tlb_valid_way
,
446 tlb_tag_way
, tlb_pte_way
, dtlb_valid_bits
,
447 dtlb_tags
, dtlb_ptes
):
449 Operates in the second cycle on the request latched in r0.req.
450 TLB updates write the entry at the end of the second cycle.
455 index
= Signal(TLB_SET_BITS
)
456 addrbits
= Signal(TLB_SET_BITS
)
459 amax
= TLB_LG_PGSZ
+ TLB_SET_BITS
461 with m
.If(m_in
.valid
):
462 comb
+= addrbits
.eq(m_in
.addr
[amin
: amax
])
464 comb
+= addrbits
.eq(d_in
.addr
[amin
: amax
])
465 comb
+= index
.eq(addrbits
)
467 # If we have any op and the previous op isn't finished,
468 # then keep the same output for next cycle.
469 with m
.If(~r0_stall
):
470 sync
+= tlb_valid_way
.eq(dtlb_valid_bits
[index
])
471 sync
+= tlb_tag_way
.eq(dtlb_tags
[index
])
472 sync
+= tlb_pte_way
.eq(dtlb_ptes
[index
])
474 def maybe_tlb_plrus(self
, m
, r1
, tlb_plru_victim
, acc
, acc_en
, lru
):
475 """Generate TLB PLRUs
480 with m
.If(TLB_NUM_WAYS
> 1):
481 for i
in range(TLB_SET_SIZE
):
483 tlb_plru
= PLRU(TLB_WAY_BITS
)
484 setattr(m
.submodules
, "maybe_plru_%d" % i
, tlb_plru
)
485 tlb_plru_acc
= Signal(TLB_WAY_BITS
)
486 tlb_plru_acc_en
= Signal()
487 tlb_plru_out
= Signal(TLB_WAY_BITS
)
489 comb
+= tlb_plru
.acc
.eq(tlb_plru_acc
)
490 comb
+= tlb_plru
.acc_en
.eq(tlb_plru_acc_en
)
491 comb
+= tlb_plru
.lru
.eq(tlb_plru_out
)
494 with m
.If(r1
.tlb_hit_index
== i
):
495 comb
+= tlb_plru
.acc_en
.eq(r1
.tlb_hit
)
497 comb
+= tlb_plru
.acc_en
.eq(0)
498 comb
+= tlb_plru
.acc
.eq(r1
.tlb_hit_way
)
500 comb
+= tlb_plru_victim
[i
].eq(tlb_plru
.lru
)
502 def tlb_search(self
, m
, tlb_req_index
, r0
, tlb_valid_way_ tlb_tag_way
,
503 tlb_pte_way
, pte
, tlb_hit
, valid_ra
, perm_attr
, ra
):
508 hitway
= Signal(TLB_WAY_BITS
)
510 eatag
= Signal(TLB_EA_TAG_BITS
)
512 TLB_LG_END
= TLB_LG_PGSZ
+ TLB_SET_BITS
513 comb
+= tlb_req_index
.eq(r0
.req
.addr
[TLB_LG_PGSZ
: TLB_LG_END
])
514 comb
+= eatag
.eq(r0
.req
.addr
[TLB_LG_END
: 64 ])
516 for i
in range(TLB_NUM_WAYS
):
517 with m
.If(tlb_valid_way(i
)
518 & read_tlb_tag(i
, tlb_tag_way
) == eatag
):
522 comb
+= tlb_hit
.eq(hit
& r0_valid
)
523 comb
+= tlb_hit_way
.eq(hitway
)
526 comb
+= pte
.eq(read_tlb_pte(hitway
, tlb_pte_way
))
529 comb
+= valid_ra
.eq(tlb_hit | ~r0
.req
.virt_mode
)
530 with m
.If(r0
.req
.virt_mode
):
531 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
532 r0
.req
.addr
[ROW_OFF_BITS
:TLB_LG_PGSZ
],
533 pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
534 comb
+= perm_attr
.eq(extract_perm_attr(pte
))
536 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
537 r0
.rq
.addr
[ROW_OFF_BITS
:REAL_ADDR_BITS
]))
539 comb
+= perm_attr
.reference
.eq(1)
540 comb
+= perm_attr
.changed
.eq(1)
541 comb
+= perm_attr
.priv
.eq(1)
542 comb
+= perm_attr
.nocache
.eq(0)
543 comb
+= perm_attr
.rd_perm
.eq(1)
544 comb
+= perm_attr
.wr_perm
.eq(1)
546 def tlb_update(self
, m
, r0_valid
, r0
, dtlb_valid_bits
, tlb_req_index
,
547 tlb_hit_way
, tlb_hit
, tlb_plru_victim
, tlb_tag_way
,
548 dtlb_tags
, tlb_pte_way
, dtlb_ptes
, dtlb_valid_bits
):
555 repl_way
= Signal(TLB_WAY_BITS
)
556 eatag
= Signal(TLB_EA_TAG_BITS
)
557 tagset
= TLBWayTags()
558 pteset
= TLBWayPtes()
560 comb
+= tlbie
.eq(r0_valid
& r0
.tlbie
)
561 comb
+= tlbwe
.eq(r0_valid
& r0
.tlbldoi
)
563 with m
.If(tlbie
& r0
.doall
):
564 # clear all valid bits at once
565 for i
in range(TLB_SET_SIZE
):
566 sync
+= dtlb_valid_bits
[i
].eq(0)
570 sync
+= dtlb_valid_bits
[tlb_req_index
][tlb_hit_way
].eq(0)
573 comb
+= repl_way
.eq(tlb_hit_way
)
575 comb
+= repl_way
.eq(tlb_plru_victim
[tlb_req_index
])
576 comb
+= eatag
.eq(r0
.req
.addr
[TLB_LG_PGSZ
+ TLB_SET_BITS
:64])
577 comb
+= tagset
.eq(tlb_tag_way
)
578 sync
+= write_tlb_tag(repl_way
, tagset
, eatag
)
579 sync
+= dtlb_tags
[tlb_req_index
].eq(tagset
)
580 comb
+= pteset
.eq(tlb_pte_way
)
581 sync
+= write_tlb_pte(repl_way
, pteset
, r0
.req
.data
)
582 sync
+= dtlb_ptes
[tlb_req_index
].eq(pteset
)
583 sync
+= dtlb_valid_bits
[tlb_req_index
][repl_way
].eq(1)
585 def maybe_plrus(self
, r1
):
591 for i
in range(NUM_LINES
):
593 plru
= PLRU(TLB_WAY_BITS
)
594 setattr(m
.submodules
, "plru%d" % i
, plru
)
595 plru_acc
= Signal(WAY_BITS
)
596 plru_acc_en
= Signal()
597 plru_out
= Signal(WAY_BITS
)
599 comb
+= plru
.acc
.eq(plru_acc
)
600 comb
+= plru
.acc_en
.eq(plru_acc_en
)
601 comb
+= plru
.lru
.eq(plru_out
)
603 with m
.If(r1
.hit_index
== i
):
604 comb
+= plru_acc_en
.eq(r1
.cache_hit
)
606 comb
+= plru_acc
.eq(r1
.hit_way
)
607 comb
+= plru_victim
[i
].eq(plru_out
)
609 def cache_tag_read(self
, m
, r0_stall
, req_index
, m_in
, d_in
,
610 cache_tag_set
, cache_tags
):
611 """Cache tag RAM read port
616 index
= Signal(INDEX_BITS
)
619 comb
+= index
.eq(req_index
)
620 with m
.Elif(m_in
.valid
):
621 comb
+= index
.eq(get_index(m_in
.addr
))
623 comb
+= index
.eq(get_index(d_in
.addr
))
624 sync
+= cache_tag_set
.eq(cache_tags
[index
])
626 def dcache_request(self
, m
, r0
, ra
, req_index
, req_row
, req_tag
,
627 r0_valid
, r1
, cache_valid_bits
, replace_way
,
628 use_forward1_next
, use_forward2_next
,
629 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
630 valid_ra
, perm_ok
, access_ok
, req_op
, req_ok
,
631 r0_stall
, m_in
, early_req_row
, d_in
):
632 """Cache request parsing and hit detection
639 hit_way
= Signal(WAY_BITS
)
645 s_tag
= Signal(TAG_BITS
)
646 s_pte
= Signal(TLB_PTE_BITS
)
647 s_ra
= Signal(REAL_ADDR_BITS
)
648 hit_set
= Signal(TLB_NUM_WAYS
)
649 hit_way_set
= HitWaySet()
650 rel_matches
= Signal(TLB_NUM_WAYS
)
653 # Extract line, row and tag from request
654 comb
+= req_index
.eq(get_index(r0
.req
.addr
))
655 comb
+= req_row
.eq(get_row(r0
.req
.addr
))
656 comb
+= req_tag
.eq(get_tag(ra
))
658 comb
+= go
.eq(r0_valid
& ~
(r0
.tlbie | r0
.tlbld
) & ~r1
.ls_error
)
660 # Test if pending request is a hit on any way
661 # In order to make timing in virtual mode,
662 # when we are using the TLB, we compare each
663 # way with each of the real addresses from each way of
664 # the TLB, and then decide later which match to use.
666 with m
.If(r0
.req
.virt_mode
):
667 comb
+= rel_matches
.eq(0)
668 for j
in range(TLB_NUM_WAYS
):
669 comb
+= s_pte
.eq(read_tlb_pte(j
, tlb_pte_way
))
670 comb
+= s_ra
.eq(Cat(r0
.req
.addr
[0:TLB_LG_PGSZ
],
671 s_pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
672 comb
+= s_tag
.eq(get_tag(s_ra
))
674 for i
in range(NUM_WAYS
):
675 with m
.If(go
& cache_valid_bits
[req_index
][i
] &
676 read_tag(i
, cache_tag_set
) == s_tag
678 comb
+= hit_way_set
[j
].eq(i
)
680 comb
+= hit_set
[j
].eq(s_hit
)
681 with m
.If(s_tag
== r1
.reload_tag
):
682 comb
+= rel_matches
[j
].eq(1)
684 comb
+= is_hit
.eq(hit_set
[tlb_hit_way
])
685 comb
+= hit_way
.eq(hit_way_set
[tlb_hit_way
])
686 comb
+= rel_match
.eq(rel_matches
[tlb_hit_way
])
688 comb
+= s_tag
.eq(get_tag(r0
.req
.addr
))
689 for i
in range(NUM_WAYS
):
690 with m
.If(go
& cache_valid_bits
[req_index
][i
] &
691 read_tag(i
, cache_tag_set
) == s_tag
):
692 comb
+= hit_way
.eq(i
)
694 with m
.If(s_tag
== r1
.reload_tag
):
695 comb
+= rel_match
.eq(1)
696 comb
+= req_same_tag
.eq(rel_match
)
698 # See if the request matches the line currently being reloaded
699 with m
.If((r1
.state
== State
.RELOAD_WAIT_ACK
) &
700 (req_index
== r1
.store_index
) & rel_match
):
701 # For a store, consider this a hit even if the row isn't
702 # valid since it will be by the time we perform the store.
703 # For a load, check the appropriate row valid bit.
704 valid
= r1
.rows_valid
[req_row
% ROW_PER_LINE
]
705 comb
+= is_hit
.eq(~r0
.req
.load | valid
)
706 comb
+= hit_way
.eq(replace_way
)
708 # Whether to use forwarded data for a load or not
709 comb
+= use_forward1_next
.eq(0)
710 with m
.If((get_row(r1
.req
.real_addr
) == req_row
)
711 & (r1
.req
.hit_way
== hit_way
))
712 # Only need to consider r1.write_bram here, since if we
713 # are writing refill data here, then we don't have a
714 # cache hit this cycle on the line being refilled.
715 # (There is the possibility that the load following the
716 # load miss that started the refill could be to the old
717 # contents of the victim line, since it is a couple of
718 # cycles after the refill starts before we see the updated
719 # cache tag. In that case we don't use the bypass.)
720 comb
+= use_forward1_next
.eq(r1
.write_bram
)
721 comb
+= use_forward2_next
.eq(0)
722 with m
.If((r1
.forward_row1
== req_row
) & (r1
.forward_way1
== hit_way
)):
723 comb
+= use_forward2_next
.eq(r1
.forward_valid1
)
725 # The way that matched on a hit
726 comb
+= req_hit_way
.eq(hit_way
)
728 # The way to replace on a miss
729 with m
.If(r1
.write_tag
):
730 replace_way
.eq(plru_victim
[r1
.store_index
])
732 comb
+= replace_way
.eq(r1
.store_way
)
734 # work out whether we have permission for this access
735 # NB we don't yet implement AMR, thus no KUAP
736 comb
+= rc_ok
.eq(perm_attr
.reference
737 & (r0
.req
.load | perm_attr
.changed
)
739 comb
+= perm_ok
.eq((r0
.req
.prive_mode | ~perm_attr
.priv
)
741 |
(r0
.req
.load
& perm_attr
.rd_perm
)
743 comb
+= access_ok
.eq(valid_ra
& perm_ok
& rc_ok
)
744 # Combine the request and cache hit status to decide what
745 # operation needs to be done
746 comb
+= nc
.eq(r0
.req
.nc | perm_attr
.nocache
)
747 comb
+= op
.eq(Op
.OP_NONE
)
749 with m
.If(~access_ok
):
750 comb
+= op
.eq(Op
.OP_BAD
)
751 with m
.Elif(cancel_store
):
752 comb
+= op
.eq(Op
.OP_STCX_FAIL
)
754 comb
+= opsel
.eq(Cat(is_hit
, nc
, r0
.req
.load
))
755 with m
.Switch(opsel
):
756 with m
.Case(Const(0b101, 3)):
757 comb
+= op
.eq(Op
.OP_LOAD_HIT
)
758 with m
.Case(Cosnt(0b100, 3)):
759 comb
+= op
.eq(Op
.OP_LOAD_MISS
)
760 with m
.Case(Const(0b110, 3)):
761 comb
+= op
.eq(Op
.OP_LOAD_NC
)
762 with m
.Case(Const(0b001, 3)):
763 comb
+= op
.eq(Op
.OP_STORE_HIT
)
764 with m
.Case(Const(0b000, 3)):
765 comb
+= op
.eq(Op
.OP_STORE_MISS
)
766 with m
.Case(Const(0b010, 3)):
767 comb
+= op
.eq(Op
.OP_STORE_MISS
)
768 with m
.Case(Const(0b011, 3)):
769 comb
+= op
.eq(Op
.OP_BAD
)
770 with m
.Case(Const(0b111, 3)):
771 comb
+= op
.eq(Op
.OP_BAD
)
773 comb
+= op
.eq(Op
.OP_NONE
)
774 comb
+= req_op
.eq(op
)
775 comb
+= req_go
.eq(go
)
777 # Version of the row number that is valid one cycle earlier
778 # in the cases where we need to read the cache data BRAM.
779 # If we're stalling then we need to keep reading the last
781 with m
.If(~r0_stall
):
782 with m
.If(m_in
.valid
):
783 comb
+= early_req_row
.eq(get_row(m_in
.addr
))
785 comb
+= early_req_row
.eq(get_row(d_in
.addr
))
787 comb
+= early_req_row
.eq(req_row
)
789 def reservation_comb(self
, m
, cancel_store
, set_rsrv
, clear_rsrv
,
790 r0_valid
, r0
, reservation
):
791 """Handle load-with-reservation and store-conditional instructions
796 with m
.If(r0_valid
& r0
.req
.reserve
):
798 # XXX generate alignment interrupt if address
799 # is not aligned XXX or if r0.req.nc = '1'
800 with m
.If(r0
.req
.load
):
801 comb
+= set_rsrv(1) # load with reservation
803 comb
+= clear_rsrv
.eq(1) # store conditional
804 with m
.If(~reservation
.valid | r0
.req
.addr
[LINE_OFF_BITS
:64]):
805 comb
+= cancel_store
.eq(1)
807 def reservation_reg(self
, m
, r0_valid
, access_ok
, clear_rsrv
,
813 with m
.If(r0_valid
& access_ok
):
814 with m
.If(clear_rsrv
):
815 sync
+= reservation
.valid
.eq(0)
816 with m
.Elif(set_rsrv
):
817 sync
+= reservation
.valid
.eq(1)
818 sync
+= reservation
.addr
.eq(r0
.req
.addr
[LINE_OFF_BITS
:64])
820 def writeback_control(self
, m
, r1
, cache_out
, d_out
, m_out
):
821 """Return data for loads & completion control logic
826 data_out
= Signal(64)
827 data_fwd
= Signal(64)
829 # Use the bypass if are reading the row that was
830 # written 1 or 2 cycles ago, including for the
831 # slow_valid = 1 case (i.e. completing a load
832 # miss or a non-cacheable load).
833 with m
.If(r1
.use_forward1
):
834 comb
+= data_fwd
.eq(r1
.forward_data1
)
836 comb
+= data_fwd
.eq(r1
.forward_data2
)
838 comb
+= data_out
.eq(cache_out
[r1
.hit_way
])
841 with m
.If(r1
.forward_sel
[i
]):
842 dsel
= data_fwd
.word_select(i
, 8)
843 comb
+= data_out
.word_select(i
, 8).eq(dsel
)
845 comb
+= d_out
.valid
.eq(r1
.ls_valid
)
846 comb
+= d_out
.data
.eq(data_out
)
847 comb
+= d_out
.store_done
.eq(~r1
.stcx_fail
)
848 comb
+= d_out
.error
.eq(r1
.ls_error
)
849 comb
+= d_out
.cache_paradox
.eq(r1
.cache_paradox
)
852 comb
+= m_out
.done
.eq(r1
.mmu_done
)
853 comb
+= m_out
.err
.eq(r1
.mmu_error
)
854 comb
+= m_out
.data
.eq(data_out
)
856 # We have a valid load or store hit or we just completed
857 # a slow op such as a load miss, a NC load or a store
859 # Note: the load hit is delayed by one cycle. However it
860 # can still not collide with r.slow_valid (well unless I
861 # miscalculated) because slow_valid can only be set on a
862 # subsequent request and not on its first cycle (the state
863 # machine must have advanced), which makes slow_valid
864 # at least 2 cycles from the previous hit_load_valid.
866 # Sanity: Only one of these must be set in any given cycle
868 if False: # TODO: need Display to get this to work
869 assert (r1
.slow_valid
& r1
.stcx_fail
) != 1 "unexpected" \
870 "slow_valid collision with stcx_fail -!- severity FAILURE"
872 assert ((r1
.slow_valid | r1
.stcx_fail
) | r1
.hit_load_valid
) != 1
873 "unexpected hit_load_delayed collision with slow_valid -!-" \
876 with m
.If(~r1
._mmu_req
):
877 # Request came from loadstore1...
878 # Load hit case is the standard path
879 with m
.If(r1
.hit_load_valid
):
880 #Display(f"completing load hit data={data_out}")
883 # error cases complete without stalling
884 with m
.If(r1
.ls_error
):
885 # Display("completing ld/st with error")
888 # Slow ops (load miss, NC, stores)
889 with m
.If(r1
.slow_valid
):
890 #Display(f"completing store or load miss data={data_out}")
894 # Request came from MMU
895 with m
.If(r1
.hit_load_valid
):
896 # Display(f"completing load hit to MMU, data={m_out.data}")
898 # error cases complete without stalling
899 with m
.If(r1
.mmu_error
):
900 #Display("combpleting MMU ld with error")
903 # Slow ops (i.e. load miss)
904 with m
.If(r1
.slow_valid
):
905 #Display("completing MMU load miss, data={m_out.data}")
910 Generate a cache RAM for each way. This handles the normal
911 reads, writes from reloads and the special store-hit update
914 Note: the BRAMs have an extra read buffer, meaning the output
915 is pipelined an extra cycle. This differs from the
916 icache. The writeback logic needs to take that into
917 account by using 1-cycle delayed signals for load hits.
921 for i
in range(NUM_WAYS
):
923 rd_addr
= Signal(ROW_BITS
)
925 wr_addr
= Signal(ROW_BITS
)
926 wr_data
= Signal(WB_DATA_BITS
)
927 wr_sel
= Signal(ROW_SIZE
)
928 wr_sel_m
= Signal(ROW_SIZE
)
929 _d_out
= Signal(WB_DATA_BITS
)
931 way
= CacheRam(ROW_BITS
, WB_DATA_BITS
, True)
932 setattr(m
.submodules
, "cacheram_%d" % i
, way
)
934 comb
+= way
.rd_en
.eq(do_read
)
935 comb
+= way
.rd_addr
.eq(rd_addr
)
936 comb
+= _d_out
.eq(way
.rd_data
)
937 comb
+= way
.wr_sel
.eq(wr_sel_m
)
938 comb
+= way
.wr_addr
.eq(wr_addr
)
939 comb
+= way
.wr_data
.eq(wr_data
)
942 comb
+= do_read
.eq(1)
943 comb
+= rd_addr
.eq(early_req_row
)
944 comb
+= cache_out
[i
].eq(_d_out
)
948 # Defaults to wishbone read responses (cache refill)
950 # For timing, the mux on wr_data/sel/addr is not
951 # dependent on anything other than the current state.
953 with m
.If(r1
.write_bram
):
954 # Write store data to BRAM. This happens one
955 # cycle after the store is in r0.
956 comb
+= wr_data
.eq(r1
.req
.data
)
957 comb
+= wr_sel
.eq(r1
.req
.byte_sel
)
958 comb
+= wr_addr
.eq(get_row(r1
.req
.real_addr
))
960 with m
.If(i
== r1
.req
.hit_way
):
961 comb
+= do_write
.eq(1)
963 # Otherwise, we might be doing a reload or a DCBZ
965 comb
+= wr_data
.eq(0)
967 comb
+= wr_data
.eq(wishbone_in
.dat
)
968 comb
+= wr_addr
.eq(r1
.store_row
)
969 comb
+= wr_sel
.eq(~
0) # all 1s
971 with m
.If((r1
.state
== State
.RELOAD_WAIT_ACK
)
972 & wishbone_in
.ack
& (relpace_way
== i
)):
973 comb
+= do_write
.eq(1)
975 # Mask write selects with do_write since BRAM
976 # doesn't have a global write-enable
978 comb
+= wr_sel_m
.eq(wr_sel
)
980 # Cache hit synchronous machine for the easy case.
981 # This handles load hits.
982 # It also handles error cases (TLB miss, cache paradox)
983 def dcache_fast_hit(self
, m
, req_op
, r0_valid
, r1
):
988 with m
.If(req_op
!= Op
.OP_NONE
):
989 #Display(f"op:{req_op} addr:{r0.req.addr} nc: {r0.req.nc}" \
990 # f"idx:{req_index} tag:{req_tag} way: {req_hit_way}"
995 sync
+= r1
.mmu_req
.eq(r0
.mmu_req
)
997 # Fast path for load/store hits.
998 # Set signals for the writeback controls.
999 sync
+= r1
.hit_way
.eq(req_hit_way
)
1000 sync
+= r1
.hit_index
.eq(req_index
)
1002 with m
.If(req_op
== Op
.OP_LOAD_HIT
):
1003 sync
+= r1
.hit_load_valid
.eq(1)
1005 sync
+= r1
.hit_load_valid
.eq(0)
1007 with m
.If((req_op
== Op
.OP_LOAD_HIT
) |
(req_op
== Op
.OP_STORE_HIT
)):
1008 sync
+= r1
.cache_hit
.eq(1)
1010 sync
+= r1
.cache_hit
.eq(0)
1012 with m
.If(req_op
== Op
.OP_BAD
):
1013 # Display(f"Signalling ld/st error valid_ra={valid_ra}"
1014 # f"rc_ok={rc_ok} perm_ok={perm_ok}"
1015 sync
+= r1
.ls_error
.eq(~r0
.mmu_req
)
1016 sync
+= r1
.mmu_error
.eq(r0
.mmu_req
)
1017 sync
+= r1
.cache_paradox
.eq(access_ok
)
1020 sync
+= r1
.ls_error
.eq(0)
1021 sync
+= r1
.mmu_error
.eq(0)
1022 sync
+= r1
.cache_paradox
.eq(0)
1024 with m
.If(req_op
== Op
.OP_STCX_FAIL
):
1027 sync
+= r1
.stcx_fail
.eq(0)
1029 # Record TLB hit information for updating TLB PLRU
1030 sync
+= r1
.tlb_hit
.eq(tlb_hit
)
1031 sync
+= r1
.tlb_hit_way
.eq(tlb_hit_way
)
1032 sync
+= r1
.tlb_hit_index
.eq(tlb_req_index
)
1034 # Memory accesses are handled by this state machine:
1036 # * Cache load miss/reload (in conjunction with "rams")
1037 # * Load hits for non-cachable forms
1038 # * Stores (the collision case is handled in "rams")
1040 # All wishbone requests generation is done here.
1041 # This machine operates at stage 1.
1042 def dcache_slow(self
, m
, r1
, use_forward1_next
, cache_valid_bits
, r0
,
1043 r0_valid
, req_op
, cache_tag
, req_go
, ra
, wb_in
):
1048 req
= MemAccessRequest()
1050 adjust_acks
= Signal(3)
1052 sync
+= r1
.use_forward1
.eq(use_forward1_next
)
1053 sync
+= r1
.forward_sel
.eq(0)
1055 with m
.If(use_forward1_next
):
1056 sync
+= r1
.forward_sel
.eq(r1
.req
.byte_sel
)
1057 with m
.Elif(use_forward2_next
):
1058 sync
+= r1
.forward_sel
.eq(r1
.forward_sel1
)
1060 sync
+= r1
.forward_data2
.eq(r1
.forward_data1
)
1061 with m
.If(r1
.write_bram
):
1062 sync
+= r1
.forward_data1
.eq(r1
.req
.data
)
1063 sync
+= r1
.forward_sel1
.eq(r1
.req
.byte_sel
)
1064 sync
+= r1
.forward_way1
.eq(r1
.req
.hit_way
)
1065 sync
+= r1
.forward_row1
.eq(get_row(r1
.req
.real_addr
))
1066 sync
+= r1
.forward_valid1
.eq(1)
1069 sync
+= r1
.forward_data1
.eq(0)
1071 sync
+= r1
.forward_data1
.eq(wb_in
.dat
)
1072 sync
+= r1
.forward_sel1
.eq(~
0) # all 1s
1073 sync
+= r1
.forward_way1
.eq(replace_way
)
1074 sync
+= r1
.forward_row1
.eq(r1
.store_row
)
1075 sync
+= r1
.forward_valid1
.eq(0)
1077 # One cycle pulses reset
1078 sync
+= r1
.slow_valid
.eq(0)
1079 sync
+= r1
.write_bram
.eq(0)
1080 sync
+= r1
.inc_acks
.eq(0)
1081 sync
+= r1
.dec_acks
.eq(0)
1083 sync
+= r1
.ls_valid
.eq(0)
1084 # complete tlbies and TLB loads in the third cycle
1085 sync
+= r1
.mmu_done
.eq(r0_valid
& (r0
.tlbie | r0
.tlbld
))
1087 with m
.If((req_op
== Op
.OP_LOAD_HIT
)
1088 |
(req_op
== Op
.OP_STCX_FAIL
)):
1089 with m
.If(~r0
.mmu_req
):
1090 sync
+= r1
.ls_valid
.eq(1)
1092 sync
+= r1
.mmu_done
.eq(1)
1094 with m
.If(r1
.write_tag
):
1095 # Store new tag in selected way
1096 for i
in range(NUM_WAYS
):
1097 with m
.If(i
== replace_way
):
1098 idx
= r1
.store_index
1099 trange
= range(i
* TAG_WIDTH
, (i
+1) * TAG_WIDTH
)
1100 sync
+= cache_tag
[idx
][trange
].eq(r1
.reload_tag
)
1101 sync
+= r1
.store_way
.eq(replace_way
)
1102 sync
+= r1
.write_tag
.eq(0)
1104 # Take request from r1.req if there is one there,
1105 # else from req_op, ra, etc.
1107 comb
+= req
.eq(r1
.req
)
1109 comb
+= req
.op
.eq(req_op
)
1110 comb
+= req
.valid
.eq(req_go
)
1111 comb
+= req
.mmu_req
.eq(r0
.mmu_req
)
1112 comb
+= req
.dcbz
.eq(r0
.req
.dcbz
)
1113 comb
+= req
.real_addr
.eq(ra
)
1115 with m
.If(~r0
.req
.dcbz
):
1116 comb
+= req
.data
.eq(r0
.req
.data
)
1118 comb
+= req
.data
.eq(0)
1120 # Select all bytes for dcbz
1121 # and for cacheable loads
1122 with m
.If(r0
.req
.dcbz |
(r0
.req
.load
& ~r0
.req
.nc
):
1123 comb
+= req
.byte_sel
.eq(~
0) # all 1s
1125 comb
+= req
.byte_sel
.eq(r0
.req
.byte_sel
)
1126 comb
+= req
.hit_way
.eq(req_hit_way
)
1127 comb
+= req
.same_tag
.eq(req_same_tag
)
1129 # Store the incoming request from r0,
1130 # if it is a slow request
1131 # Note that r1.full = 1 implies req_op = OP_NONE
1132 with m
.If((req_op
== Op
.OP_LOAD_MISS
)
1133 |
(req_op
== Op
.OP_LOAD_NC
)
1134 |
(req_op
== Op
.OP_STORE_MISS
)
1135 |
(req_op
== Op
.OP_STORE_HIT
)):
1137 sync
+= r1
.full
.eq(1)
1139 # Main state machine
1140 with m
.Switch(r1
.state
):
1142 with m
.Case(State
.IDLE
)
1143 # XXX check 'left downto. probably means len(r1.wb.adr)
1144 # r1.wb.adr <= req.real_addr(
1145 # r1.wb.adr'left downto 0
1147 sync
+= r1
.wb
.adr
.eq(req
.real_addr
[0:r1
.wb
.adr
])
1148 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1149 sync
+= r1
.wb
.dat
.eq(req
.data
)
1150 sync
+= r1
.dcbz
.eq(req
.dcbz
)
1152 # Keep track of our index and way
1153 # for subsequent stores.
1154 sync
+= r1
.store_index
.eq(get_index(req
.real_addr
))
1155 sync
+= r1
.store_row
.eq(get_row(req
.real_addr
))
1156 sync
+= r1
.end_row_ix
.eq(
1157 get_row_of_line(get_row(req
.real_addr
))
1159 sync
+= r1
.reload_tag
.eq(get_tag(req
.real_addr
))
1160 sync
+= r1
.req
.same_tag
.eq(1)
1162 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1163 sync
+= r1
.store_way
.eq(req
.hit_way
)
1165 # Reset per-row valid bits,
1166 # ready for handling OP_LOAD_MISS
1167 for i
in range(ROW_PER_LINE
):
1168 sync
+= r1
.rows_valid
[i
].eq(0)
1170 with m
.Switch(req
.op
):
1171 with m
.Case(Op
.OP_LOAD_HIT
):
1172 # stay in IDLE state
1175 with m
.Case(Op
.OP_LOAD_MISS
):
1176 #Display(f"cache miss real addr:" \
1177 # f"{req_real_addr}" \
1178 # f" idx:{get_index(req_real_addr)}" \
1179 # f" tag:{get_tag(req.real_addr)}")
1182 # Start the wishbone cycle
1183 sync
+= r1
.wb
.we
.eq(0)
1184 sync
+= r1
.wb
.cyc
.eq(1)
1185 sync
+= r1
.wb
.stb
.eq(1)
1187 # Track that we had one request sent
1188 sync
+= r1
.state
.eq(State
.RELOAD_WAIT_ACK
)
1189 sync
+= r1
.write_tag
.eq(1)
1191 with m
.Case(Op
.OP_LOAD_NC
):
1192 sync
+= r1
.wb
.cyc
.eq(1)
1193 sync
+= r1
.wb
.stb
.eq(1)
1194 sync
+= r1
.wb
.we
.eq(0)
1195 sync
+= r1
.state
.eq(State
.NC_LOAD_WAIT_ACK
)
1197 with m
.Case(Op
.OP_STORE_HIT
, Op
.OP_STORE_MISS
):
1198 with m
.If(~req
.bcbz
):
1199 sync
+= r1
.state
.eq(State
.STORE_WAIT_ACK
)
1200 sync
+= r1
.acks_pending
.eq(1)
1201 sync
+= r1
.full
.eq(0)
1202 sync
+= r1
.slow_valid
.eq(1)
1204 with m
.If(~req
.mmu_req
):
1205 sync
+= r1
.ls_valid
.eq(1)
1207 sync
+= r1
.mmu_done
.eq(1)
1209 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1210 sync
+= r1
.write_bram
.eq(1)
1212 sync
+= r1
.state
.eq(Op
.RELOAD_WAIT_ACK
)
1214 with m
.If(req
.op
== Op
.OP_STORE_MISS
):
1215 sync
+= r1
.write_tag
.eq(1)
1217 sync
+= r1
.wb
.we
.eq(1)
1218 sync
+= r1
.wb
.cyc
.eq(1)
1219 sync
+= r1
.wb
.stb
.eq(1)
1221 # OP_NONE and OP_BAD do nothing
1222 # OP_BAD & OP_STCX_FAIL were
1223 # handled above already
1224 with m
.Case(Op
.OP_NONE
):
1226 with m
.Case(OP_BAD
):
1228 with m
.Case(OP_STCX_FAIL
):
1231 with m
.Case(State
.RELOAD_WAIT_ACK
):
1232 # Requests are all sent if stb is 0
1233 comb
+= stbs_done
.eq(~r1
.wb
.stb
)
1235 with m
.If(~wb_in
.stall
& ~stbs_done
):
1236 # That was the last word?
1237 # We are done sending.
1238 # Clear stb and set stbs_done
1239 # so we can handle an eventual
1240 # last ack on the same cycle.
1241 with m
.If(is_last_row_addr(
1242 r1
.wb
.adr
, r1
.end_row_ix
)):
1243 sync
+= r1
.wb
.stb
.eq(0)
1244 comb
+= stbs_done
.eq(0)
1246 # Calculate the next row address
1247 sync
+= r1
.wb
.adr
.eq(next_row_addr(r1
.wb
.adr
))
1249 # Incoming acks processing
1250 sync
+= r1
.forward_valid1
.eq(wb_in
.ack
)
1251 with m
.If(wb_in
.ack
):
1252 # XXX needs an Array bit-accessor here
1253 sync
+= r1
.rows_valid
[r1
.store_row
% ROW_PER_LINE
].eq(1)
1255 # If this is the data we were looking for,
1256 # we can complete the request next cycle.
1257 # Compare the whole address in case the
1258 # request in r1.req is not the one that
1259 # started this refill.
1260 with m
.If(r1
.full
& r1
.req
.same_tag
&
1261 ((r1
.dcbz
& r1
.req
.dcbz
) |
1262 (~r1
.dcbz
& (r1
.req
.op
== Op
.OP_LOAD_MISS
))) &
1263 (r1
.store_row
== get_row(r1
.req
.real_addr
))):
1264 sync
+= r1
.full
.eq(0)
1265 sync
+= r1
.slow_valid
.eq(1)
1266 with m
.If(~r1
.mmu_req
):
1267 sync
+= r1
.ls_valid
.eq(1)
1269 sync
+= r1
.mmu_done
.eq(1)
1270 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1271 sync
+= r1
.use_forward1
.eq(1)
1273 # Check for completion
1274 with m
.If(stbs_done
& is_last_row(r1
.store_row
,
1276 # Complete wishbone cycle
1277 sync
+= r1
.wb
.cyc
.eq(0)
1279 # Cache line is now valid
1280 cv
= cache_valid_bits
[r1
.store_index
]
1281 sync
+= cv
[r1
.store_way
].eq(1)
1282 sync
+= r1
.state
.eq(State
.IDLE
)
1284 # Increment store row counter
1285 sync
+= r1
.store_row
.eq(next_row(r1
.store_row
))
1287 with m
.Case(State
.STORE_WAIT_ACK
):
1288 comb
+= stbs_done
.eq(~r1
.wb
.stb
)
1289 comb
+= acks
.eq(r1
.acks_pending
)
1291 with m
.If(r1
.inc_acks
!= r1
.dec_acks
):
1292 with m
.If(r1
.inc_acks
):
1293 comb
+= adjust_acks
.eq(acks
+ 1)
1295 comb
+= adjust_acks
.eq(acks
- 1)
1297 comb
+= adjust_acks
.eq(acks
)
1299 sync
+= r1
.acks_pending
.eq(adjust_acks
)
1301 # Clear stb when slave accepted request
1302 with m
.If(~wb_in
.stall
):
1303 # See if there is another store waiting
1304 # to be done which is in the same real page.
1305 with m
.If(req
.valid
):
1306 ra
= req
.real_addr
[0:SET_SIZE_BITS
]
1307 sync
+= r1
.wb
.adr
[0:SET_SIZE_BITS
].eq(ra
)
1308 sync
+= r1
.wb
.dat
.eq(req
.data
)
1309 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1311 with m
.Elif((adjust_acks
< 7) & req
.same_tag
&
1312 ((req
.op
== Op
.Op_STORE_MISS
)
1313 |
(req
.op
== Op
.OP_SOTRE_HIT
))):
1314 sync
+= r1
.wb
.stb
.eq(1)
1315 comb
+= stbs_done
.eq(0)
1317 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1318 sync
+= r1
.write_bram
.eq(1)
1319 sync
+= r1
.full
.eq(0)
1320 sync
+= r1
.slow_valid
.eq(1)
1322 # Store requests never come from the MMU
1323 sync
+= r1
.ls_valid
.eq(1)
1324 comb
+= stbs_done
.eq(0)
1325 sync
+= r1
.inc_acks
.eq(1)
1327 sync
+= r1
.wb
.stb
.eq(0)
1328 comb
+= stbs_done
.eq(1)
1330 # Got ack ? See if complete.
1331 with m
.If(wb_in
.ack
):
1332 with m
.If(stbs_done
& (adjust_acks
== 1))
1333 sync
+= r1
.state
.eq(State
.IDLE
)
1334 sync
+= r1
.wb
.cyc
.eq(0)
1335 sync
+= r1
.wb
.stb
.eq(0)
1336 sync
+= r1
.dec_acks
.eq(1)
1338 with m
.Case(State
.NC_LOAD_WAIT_ACK
):
1339 # Clear stb when slave accepted request
1340 with m
.If(~wb_in
.stall
):
1341 sync
+= r1
.wb
.stb
.eq(0)
1343 # Got ack ? complete.
1344 with m
.If(wb_in
.ack
):
1345 sync
+= r1
.state
.eq(State
.IDLE
)
1346 sync
+= r1
.full
.eq(0)
1347 sync
+= r1
.slow_valid
.eq(1)
1349 with m
.If(~r1
.mmu_req
):
1350 sync
+= r1
.ls_valid
.eq(1)
1352 sync
+= r1
.mmu_done
.eq(1)
1354 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1355 sync
+= r1
.use_forward1
.eq(1)
1356 sync
+= r1
.wb
.cyc
.eq(0)
1357 sync
+= r1
.wb
.stb
.eq(0)
1359 def dcache_log(self
, m
, r1
, valid_ra
, tlb_hit_way
, stall_out
,
1360 d_out
, wb_in
, log_out
):
1364 sync
+= log_out
.eq(Cat(r1
.state
[:3], valid_ra
, tlb_hit_way
[:3],
1365 stall_out
, req_op
[:3], d_out
.valid
, d_out
.error
,
1366 r1
.wb
.cyc
, r1
.wb
.stb
, wb_in
.ack
, wb_in
.stall
,
1369 def elaborate(self
, platform
):
1371 # Storage. Hopefully "cache_rows" is a BRAM, the rest is LUTs
1372 cache_tags
= CacheTagArray()
1373 cache_tag_set
= Signal(TAG_RAM_WIDTH
)
1374 cache_valid_bits
= CacheValidBitsArray()
1376 # TODO attribute ram_style : string;
1377 # TODO attribute ram_style of cache_tags : signal is "distributed";
1379 """note: these are passed to nmigen.hdl.Memory as "attributes".
1380 don't know how, just that they are.
1382 dtlb_valid_bits
= TLBValidBitsArray()
1383 dtlb_tags
= TLBTagsArray()
1384 dtlb_ptes
= TLBPtesArray()
1385 # TODO attribute ram_style of
1386 # dtlb_tags : signal is "distributed";
1387 # TODO attribute ram_style of
1388 # dtlb_ptes : signal is "distributed";
1395 reservation
= Reservation()
1397 # Async signals on incoming request
1398 req_index
= Signal(INDEX_BITS
)
1399 req_row
= Signal(ROW_BITS
)
1400 req_hit_way
= Signal(WAY_BITS
)
1401 req_tag
= Signal(TAG_BITS
)
1403 req_data
= Signal(64)
1404 req_same_tag
= Signal()
1407 early_req_row
= Signal(ROW_BITS
)
1409 cancel_store
= Signal()
1411 clear_rsrv
= Signal()
1416 use_forward1_next
= Signal()
1417 use_forward2_next
= Signal()
1419 cache_out
= CacheRamOut()
1421 plru_victim
= PLRUOut()
1422 replace_way
= Signal(WAY_BITS
)
1424 # Wishbone read/write/cache write formatting signals
1428 tlb_tag_way
= Signal(TLB_TAG_WAY_BITS
)
1429 tlb_pte_way
= Signal(TLB_PTE_WAY_BITS
)
1430 tlb_valid_way
= Signal(TLB_NUM_WAYS
)
1431 tlb_req_index
= Signal(TLB_SET_BITS
)
1433 tlb_hit_way
= Signal(TLB_WAY_BITS
)
1434 pte
= Signal(TLB_PTE_BITS
)
1435 ra
= Signal(REAL_ADDR_BITS
)
1437 perm_attr
= PermAttr()
1440 access_ok
= Signal()
1442 tlb_plru_victim
= TLBPLRUOut()
1444 # we don't yet handle collisions between loadstore1 requests
1446 comb
+= m_out
.stall
.eq(0)
1448 # Hold off the request in r0 when r1 has an uncompleted request
1449 comb
+= r0_stall
.eq(r0_full
& r1
.full
)
1450 comb
+= r0_valid
.eq(r0_full
& ~r1
.full
)
1451 comb
+= stall_out
.eq(r0_stall
)
1453 # Wire up wishbone request latch out of stage 1
1454 comb
+= wishbone_out
.eq(r1
.wb
)
1460 # entity dcache_tb is
1463 # architecture behave of dcache_tb is
1464 # signal clk : std_ulogic;
1465 # signal rst : std_ulogic;
1467 # signal d_in : Loadstore1ToDcacheType;
1468 # signal d_out : DcacheToLoadstore1Type;
1470 # signal m_in : MmuToDcacheType;
1471 # signal m_out : DcacheToMmuType;
1473 # signal wb_bram_in : wishbone_master_out;
1474 # signal wb_bram_out : wishbone_slave_out;
1476 # constant clk_period : time := 10 ns;
1478 # dcache0: entity work.dcache
1491 # wishbone_out => wb_bram_in,
1492 # wishbone_in => wb_bram_out
1495 # -- BRAM Memory slave
1496 # bram0: entity work.wishbone_bram_wrapper
1498 # MEMORY_SIZE => 1024,
1499 # RAM_INIT_FILE => "icache_test.bin"
1504 # wishbone_in => wb_bram_in,
1505 # wishbone_out => wb_bram_out
1508 # clk_process: process
1511 # wait for clk_period/2;
1513 # wait for clk_period/2;
1516 # rst_process: process
1519 # wait for 2*clk_period;
1527 # d_in.valid <= '0';
1530 # d_in.addr <= (others => '0');
1531 # d_in.data <= (others => '0');
1532 # m_in.valid <= '0';
1533 # m_in.addr <= (others => '0');
1534 # m_in.pte <= (others => '0');
1536 # wait for 4*clk_period;
1537 # wait until rising_edge(clk);
1539 # -- Cacheable read of address 4
1542 # d_in.addr <= x"0000000000000004";
1543 # d_in.valid <= '1';
1544 # wait until rising_edge(clk);
1545 # d_in.valid <= '0';
1547 # wait until rising_edge(clk) and d_out.valid = '1';
1548 # assert d_out.data = x"0000000100000000"
1549 # report "data @" & to_hstring(d_in.addr) &
1550 # "=" & to_hstring(d_out.data) &
1551 # " expected 0000000100000000"
1553 # -- wait for clk_period;
1555 # -- Cacheable read of address 30
1558 # d_in.addr <= x"0000000000000030";
1559 # d_in.valid <= '1';
1560 # wait until rising_edge(clk);
1561 # d_in.valid <= '0';
1563 # wait until rising_edge(clk) and d_out.valid = '1';
1564 # assert d_out.data = x"0000000D0000000C"
1565 # report "data @" & to_hstring(d_in.addr) &
1566 # "=" & to_hstring(d_out.data) &
1567 # " expected 0000000D0000000C"
1570 # -- Non-cacheable read of address 100
1573 # d_in.addr <= x"0000000000000100";
1574 # d_in.valid <= '1';
1575 # wait until rising_edge(clk);
1576 # d_in.valid <= '0';
1577 # wait until rising_edge(clk) and d_out.valid = '1';
1578 # assert d_out.data = x"0000004100000040"
1579 # report "data @" & to_hstring(d_in.addr) &
1580 # "=" & to_hstring(d_out.data) &
1581 # " expected 0000004100000040"
1584 # wait until rising_edge(clk);
1585 # wait until rising_edge(clk);
1586 # wait until rising_edge(clk);
1587 # wait until rising_edge(clk);
1592 def dcache_sim(dut
):
1594 yield dut
.d_in
.valid
.eq(0)
1595 yield dut
.d_in
.load
.eq(0)
1596 yield dut
.d_in
.nc
.eq(0)
1597 yield dut
.d_in
.adrr
.eq(0)
1598 yield dut
.d_in
.data
.eq(0)
1599 yield dut
.m_in
.valid
.eq(0)
1600 yield dut
.m_in
.addr
.eq(0)
1601 yield dut
.m_in
.pte
.eq(0)
1602 # wait 4 * clk_period
1607 # wait_until rising_edge(clk)
1609 # Cacheable read of address 4
1610 yield dut
.d_in
.load
.eq(1)
1611 yield dut
.d_in
.nc
.eq(0)
1612 yield dut
.d_in
.addr
.eq(Const(0x0000000000000004, 64))
1613 yield dut
.d_in
.valid
.eq(1)
1614 # wait-until rising_edge(clk)
1616 yield dut
.d_in
.valid
.eq(0)
1618 while not (yield dut
.d_out
.valid
):
1620 assert dut
.d_out
.data
== Const(0x0000000100000000, 64) f
"data @" \
1621 f
"{dut.d_in.addr}={dut.d_in.data} expected 0000000100000000" \
1622 " -!- severity failure"
1625 # Cacheable read of address 30
1626 yield dut
.d_in
.load
.eq(1)
1627 yield dut
.d_in
.nc
.eq(0)
1628 yield dut
.d_in
.addr
.eq(Const(0x0000000000000030, 64))
1629 yield dut
.d_in
.valid
.eq(1)
1631 yield dut
.d_in
.valid
.eq(0)
1633 while not (yield dut
.d_out
.valid
):
1635 assert dut
.d_out
.data
== Const(0x0000000D0000000C, 64) f
"data @" \
1636 f
"{dut.d_in.addr}={dut.d_out.data} expected 0000000D0000000C" \
1637 f
"-!- severity failure"
1639 # Non-cacheable read of address 100
1640 yield dut
.d_in
.load
.eq(1)
1641 yield dut
.d_in
.nc
.eq(1)
1642 yield dut
.d_in
.addr
.eq(Const(0x0000000000000100, 64))
1643 yield dut
.d_in
.valid
.eq(1)
1645 yield dut
.d_in
.valid
.eq(0)
1647 while not (yield dut
.d_out
.valid
):
1649 assert dut
.d_out
.data
== Const(0x0000004100000040, 64) f
"data @" \
1650 f
"{dut.d_in.addr}={dut.d_out.data} expected 0000004100000040" \
1651 f
"-!- severity failure"
1661 vl
= rtlil
.convert(dut
, ports
=[])
1662 with
open("test_dcache.il", "w") as f
:
1665 run_simulation(dut
, dcache_sim(), vcd_name
='test_dcache.vcd')
1667 if __name__
== '__main__':