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
9 IMPORTANT: for store, the data is sampled the cycle AFTER the "valid"
14 * https://libre-soc.org/3d_gpu/architecture/set_associative_cache.jpg
15 * https://bugs.libre-soc.org/show_bug.cgi?id=469
21 from nmutil
.gtkw
import write_gtkw
23 sys
.setrecursionlimit(1000000)
25 from enum
import Enum
, unique
27 from nmigen
import (Module
, Signal
, Elaboratable
, Cat
, Repl
, Array
, Const
,
29 from nmutil
.util
import Display
30 from nmigen
.lib
.coding
import Decoder
32 from copy
import deepcopy
33 from random
import randint
, seed
35 from nmigen_soc
.wishbone
.bus
import Interface
37 from nmigen
.cli
import main
38 from nmutil
.iocontrol
import RecordObject
39 from nmigen
.utils
import log2_int
40 from soc
.experiment
.mem_types
import (LoadStore1ToDCacheType
,
41 DCacheToLoadStore1Type
,
45 from soc
.experiment
.wb_types
import (WB_ADDR_BITS
, WB_DATA_BITS
, WB_SEL_BITS
,
46 WBAddrType
, WBDataType
, WBSelType
,
47 WBMasterOut
, WBSlaveOut
,
48 WBMasterOutVector
, WBSlaveOutVector
,
49 WBIOMasterOut
, WBIOSlaveOut
)
51 from soc
.experiment
.cache_ram
import CacheRam
52 #from soc.experiment.plru import PLRU
53 from nmutil
.plru
import PLRU
, PLRUs
56 from soc
.bus
.sram
import SRAM
57 from nmigen
import Memory
58 from nmigen
.cli
import rtlil
60 # NOTE: to use cxxsim, export NMIGEN_SIM_MODE=cxxsim from the shell
61 # Also, check out the cxxsim nmigen branch, and latest yosys from git
62 from nmutil
.sim_tmp_alternative
import Simulator
64 from nmutil
.util
import wrap
67 # TODO: make these parameters of DCache at some point
68 LINE_SIZE
= 64 # Line size in bytes
69 NUM_LINES
= 16 # Number of lines in a set
70 NUM_WAYS
= 4 # Number of ways
71 TLB_SET_SIZE
= 64 # L1 DTLB entries per set
72 TLB_NUM_WAYS
= 2 # L1 DTLB number of sets
73 TLB_LG_PGSZ
= 12 # L1 DTLB log_2(page_size)
74 LOG_LENGTH
= 0 # Non-zero to enable log data collection
76 # BRAM organisation: We never access more than
77 # -- WB_DATA_BITS at a time so to save
78 # -- resources we make the array only that wide, and
79 # -- use consecutive indices to make a cache "line"
81 # -- ROW_SIZE is the width in bytes of the BRAM
82 # -- (based on WB, so 64-bits)
83 ROW_SIZE
= WB_DATA_BITS
// 8;
85 # ROW_PER_LINE is the number of row (wishbone
86 # transactions) in a line
87 ROW_PER_LINE
= LINE_SIZE
// ROW_SIZE
89 # BRAM_ROWS is the number of rows in BRAM needed
90 # to represent the full dcache
91 BRAM_ROWS
= NUM_LINES
* ROW_PER_LINE
93 print ("ROW_SIZE", ROW_SIZE
)
94 print ("ROW_PER_LINE", ROW_PER_LINE
)
95 print ("BRAM_ROWS", BRAM_ROWS
)
96 print ("NUM_WAYS", NUM_WAYS
)
98 # Bit fields counts in the address
100 # REAL_ADDR_BITS is the number of real address
104 # ROW_BITS is the number of bits to select a row
105 ROW_BITS
= log2_int(BRAM_ROWS
)
107 # ROW_LINE_BITS is the number of bits to select
108 # a row within a line
109 ROW_LINE_BITS
= log2_int(ROW_PER_LINE
)
111 # LINE_OFF_BITS is the number of bits for
112 # the offset in a cache line
113 LINE_OFF_BITS
= log2_int(LINE_SIZE
)
115 # ROW_OFF_BITS is the number of bits for
116 # the offset in a row
117 ROW_OFF_BITS
= log2_int(ROW_SIZE
)
119 # INDEX_BITS is the number if bits to
120 # select a cache line
121 INDEX_BITS
= log2_int(NUM_LINES
)
123 # SET_SIZE_BITS is the log base 2 of the set size
124 SET_SIZE_BITS
= LINE_OFF_BITS
+ INDEX_BITS
126 # TAG_BITS is the number of bits of
127 # the tag part of the address
128 TAG_BITS
= REAL_ADDR_BITS
- SET_SIZE_BITS
130 # TAG_WIDTH is the width in bits of each way of the tag RAM
131 TAG_WIDTH
= TAG_BITS
+ 7 - ((TAG_BITS
+ 7) % 8)
133 # WAY_BITS is the number of bits to select a way
134 WAY_BITS
= log2_int(NUM_WAYS
)
136 # Example of layout for 32 lines of 64 bytes:
138 .. tag |index| line |
140 .. | |---| | ROW_LINE_BITS (3)
141 .. | |--- - --| LINE_OFF_BITS (6)
142 .. | |- --| ROW_OFF_BITS (3)
143 .. |----- ---| | ROW_BITS (8)
144 .. |-----| | INDEX_BITS (5)
145 .. --------| | TAG_BITS (45)
148 print ("Dcache TAG %d IDX %d ROW_BITS %d ROFF %d LOFF %d RLB %d" % \
149 (TAG_BITS
, INDEX_BITS
, ROW_BITS
,
150 ROW_OFF_BITS
, LINE_OFF_BITS
, ROW_LINE_BITS
))
151 print ("index @: %d-%d" % (LINE_OFF_BITS
, SET_SIZE_BITS
))
152 print ("row @: %d-%d" % (LINE_OFF_BITS
, ROW_OFF_BITS
))
153 print ("tag @: %d-%d width %d" % (SET_SIZE_BITS
, REAL_ADDR_BITS
, TAG_WIDTH
))
155 TAG_RAM_WIDTH
= TAG_WIDTH
* NUM_WAYS
157 print ("TAG_RAM_WIDTH", TAG_RAM_WIDTH
)
158 print (" TAG_WIDTH", TAG_WIDTH
)
159 print (" NUM_WAYS", NUM_WAYS
)
162 tag_layout
= [('valid', 1),
163 ('tag', TAG_RAM_WIDTH
),
165 return Array(Record(tag_layout
, name
="tag%d" % x
) for x
in range(NUM_LINES
))
167 def RowPerLineValidArray():
168 return Array(Signal(name
="rows_valid%d" % x
) \
169 for x
in range(ROW_PER_LINE
))
172 TLB_SET_BITS
= log2_int(TLB_SET_SIZE
)
173 TLB_WAY_BITS
= log2_int(TLB_NUM_WAYS
)
174 TLB_EA_TAG_BITS
= 64 - (TLB_LG_PGSZ
+ TLB_SET_BITS
)
175 TLB_TAG_WAY_BITS
= TLB_NUM_WAYS
* TLB_EA_TAG_BITS
177 TLB_PTE_WAY_BITS
= TLB_NUM_WAYS
* TLB_PTE_BITS
;
180 return (1<<log2_int(x
, False)) == x
182 assert (LINE_SIZE
% ROW_SIZE
) == 0, "LINE_SIZE not multiple of ROW_SIZE"
183 assert ispow2(LINE_SIZE
), "LINE_SIZE not power of 2"
184 assert ispow2(NUM_LINES
), "NUM_LINES not power of 2"
185 assert ispow2(ROW_PER_LINE
), "ROW_PER_LINE not power of 2"
186 assert ROW_BITS
== (INDEX_BITS
+ ROW_LINE_BITS
), "geometry bits don't add up"
187 assert (LINE_OFF_BITS
== ROW_OFF_BITS
+ ROW_LINE_BITS
), \
188 "geometry bits don't add up"
189 assert REAL_ADDR_BITS
== (TAG_BITS
+ INDEX_BITS
+ LINE_OFF_BITS
), \
190 "geometry bits don't add up"
191 assert REAL_ADDR_BITS
== (TAG_BITS
+ ROW_BITS
+ ROW_OFF_BITS
), \
192 "geometry bits don't add up"
193 assert 64 == WB_DATA_BITS
, "Can't yet handle wb width that isn't 64-bits"
194 assert SET_SIZE_BITS
<= TLB_LG_PGSZ
, "Set indexed by virtual address"
198 return Record([('valid', 1),
199 ('way', TLB_WAY_BITS
)], name
=name
)
202 return Array(Signal(TLB_EA_TAG_BITS
, name
="tlbtagea%d" % x
) \
203 for x
in range (TLB_NUM_WAYS
))
206 tlb_layout
= [('valid', TLB_NUM_WAYS
),
207 ('tag', TLB_TAG_WAY_BITS
),
208 ('pte', TLB_PTE_WAY_BITS
)
210 return Record(tlb_layout
, name
=name
)
213 return Array(Signal(TLB_NUM_WAYS
, name
="tlb_valid%d" % x
)
214 for x
in range(TLB_SET_SIZE
))
217 return Array(Signal(WAY_BITS
, name
="hitway_%d" % x
) \
218 for x
in range(TLB_NUM_WAYS
))
220 # Cache RAM interface
222 return Array(Signal(WB_DATA_BITS
, name
="cache_out%d" % x
) \
223 for x
in range(NUM_WAYS
))
225 # PLRU output interface
227 return Array(Signal(WAY_BITS
, name
="plru_out%d" % x
) \
228 for x
in range(NUM_LINES
))
230 # TLB PLRU output interface
232 return Array(Signal(TLB_WAY_BITS
, name
="tlbplru_out%d" % x
) \
233 for x
in range(TLB_SET_SIZE
))
235 # Helper functions to decode incoming requests
237 # Return the cache line index (tag index) for an address
239 return addr
[LINE_OFF_BITS
:SET_SIZE_BITS
]
241 # Return the cache row index (data memory) for an address
243 return addr
[ROW_OFF_BITS
:SET_SIZE_BITS
]
245 # Return the index of a row within a line
246 def get_row_of_line(row
):
247 return row
[:ROW_BITS
][:ROW_LINE_BITS
]
249 # Returns whether this is the last row of a line
250 def is_last_row_addr(addr
, last
):
251 return addr
[ROW_OFF_BITS
:LINE_OFF_BITS
] == last
253 # Returns whether this is the last row of a line
254 def is_last_row(row
, last
):
255 return get_row_of_line(row
) == last
257 # Return the next row in the current cache line. We use a
258 # dedicated function in order to limit the size of the
259 # generated adder to be only the bits within a cache line
260 # (3 bits with default settings)
262 row_v
= row
[0:ROW_LINE_BITS
] + 1
263 return Cat(row_v
[:ROW_LINE_BITS
], row
[ROW_LINE_BITS
:])
265 # Get the tag value from the address
267 return addr
[SET_SIZE_BITS
:REAL_ADDR_BITS
]
269 # Read a tag from a tag memory row
270 def read_tag(way
, tagset
):
271 return tagset
.word_select(way
, TAG_WIDTH
)[:TAG_BITS
]
273 # Read a TLB tag from a TLB tag memory row
274 def read_tlb_tag(way
, tags
):
275 return tags
.word_select(way
, TLB_EA_TAG_BITS
)
277 # Write a TLB tag to a TLB tag memory row
278 def write_tlb_tag(way
, tags
, tag
):
279 return read_tlb_tag(way
, tags
).eq(tag
)
281 # Read a PTE from a TLB PTE memory row
282 def read_tlb_pte(way
, ptes
):
283 return ptes
.word_select(way
, TLB_PTE_BITS
)
285 def write_tlb_pte(way
, ptes
, newpte
):
286 return read_tlb_pte(way
, ptes
).eq(newpte
)
289 # Record for storing permission, attribute, etc. bits from a PTE
290 class PermAttr(RecordObject
):
291 def __init__(self
, name
=None):
292 super().__init
__(name
=name
)
293 self
.reference
= Signal()
294 self
.changed
= Signal()
295 self
.nocache
= Signal()
297 self
.rd_perm
= Signal()
298 self
.wr_perm
= Signal()
301 def extract_perm_attr(pte
):
306 # Type of operation on a "valid" input
310 OP_BAD
= 1 # NC cache hit, TLB miss, prot/RC failure
311 OP_STCX_FAIL
= 2 # conditional store w/o reservation
312 OP_LOAD_HIT
= 3 # Cache hit on load
313 OP_LOAD_MISS
= 4 # Load missing cache
314 OP_LOAD_NC
= 5 # Non-cachable load
315 OP_STORE_HIT
= 6 # Store hitting cache
316 OP_STORE_MISS
= 7 # Store missing cache
319 # Cache state machine
322 IDLE
= 0 # Normal load hit processing
323 RELOAD_WAIT_ACK
= 1 # Cache reload wait ack
324 STORE_WAIT_ACK
= 2 # Store wait ack
325 NC_LOAD_WAIT_ACK
= 3 # Non-cachable load wait ack
330 # In order to make timing, we use the BRAMs with
331 # an output buffer, which means that the BRAM
332 # output is delayed by an extra cycle.
334 # Thus, the dcache has a 2-stage internal pipeline
335 # for cache hits with no stalls.
337 # All other operations are handled via stalling
338 # in the first stage.
340 # The second stage can thus complete a hit at the same
341 # time as the first stage emits a stall for a complex op.
343 # Stage 0 register, basically contains just the latched request
345 class RegStage0(RecordObject
):
346 def __init__(self
, name
=None):
347 super().__init
__(name
=name
)
348 self
.req
= LoadStore1ToDCacheType(name
="lsmem")
349 self
.tlbie
= Signal() # indicates a tlbie request (from MMU)
350 self
.doall
= Signal() # with tlbie, indicates flush whole TLB
351 self
.tlbld
= Signal() # indicates a TLB load request (from MMU)
352 self
.mmu_req
= Signal() # indicates source of request
353 self
.d_valid
= Signal() # indicates req.data is valid now
356 class MemAccessRequest(RecordObject
):
357 def __init__(self
, name
=None):
358 super().__init
__(name
=name
)
360 self
.valid
= Signal()
362 self
.real_addr
= Signal(REAL_ADDR_BITS
)
363 self
.data
= Signal(64)
364 self
.byte_sel
= Signal(8)
365 self
.hit_way
= Signal(WAY_BITS
)
366 self
.same_tag
= Signal()
367 self
.mmu_req
= Signal()
370 # First stage register, contains state for stage 1 of load hits
371 # and for the state machine used by all other operations
372 class RegStage1(RecordObject
):
373 def __init__(self
, name
=None):
374 super().__init
__(name
=name
)
375 # Info about the request
376 self
.full
= Signal() # have uncompleted request
377 self
.mmu_req
= Signal() # request is from MMU
378 self
.req
= MemAccessRequest(name
="reqmem")
381 self
.hit_way
= Signal(WAY_BITS
)
382 self
.hit_load_valid
= Signal()
383 self
.hit_index
= Signal(INDEX_BITS
)
384 self
.cache_hit
= Signal()
387 self
.tlb_hit
= TLBHit("tlb_hit")
388 self
.tlb_hit_index
= Signal(TLB_SET_BITS
)
390 # 2-stage data buffer for data forwarded from writes to reads
391 self
.forward_data1
= Signal(64)
392 self
.forward_data2
= Signal(64)
393 self
.forward_sel1
= Signal(8)
394 self
.forward_valid1
= Signal()
395 self
.forward_way1
= Signal(WAY_BITS
)
396 self
.forward_row1
= Signal(ROW_BITS
)
397 self
.use_forward1
= Signal()
398 self
.forward_sel
= Signal(8)
400 # Cache miss state (reload state machine)
401 self
.state
= Signal(State
)
403 self
.write_bram
= Signal()
404 self
.write_tag
= Signal()
405 self
.slow_valid
= Signal()
406 self
.wb
= WBMasterOut("wb")
407 self
.reload_tag
= Signal(TAG_BITS
)
408 self
.store_way
= Signal(WAY_BITS
)
409 self
.store_row
= Signal(ROW_BITS
)
410 self
.store_index
= Signal(INDEX_BITS
)
411 self
.end_row_ix
= Signal(ROW_LINE_BITS
)
412 self
.rows_valid
= RowPerLineValidArray()
413 self
.acks_pending
= Signal(3)
414 self
.inc_acks
= Signal()
415 self
.dec_acks
= Signal()
417 # Signals to complete (possibly with error)
418 self
.ls_valid
= Signal()
419 self
.ls_error
= Signal()
420 self
.mmu_done
= Signal()
421 self
.mmu_error
= Signal()
422 self
.cache_paradox
= Signal()
424 # Signal to complete a failed stcx.
425 self
.stcx_fail
= Signal()
428 # Reservation information
429 class Reservation(RecordObject
):
432 self
.valid
= Signal()
433 self
.addr
= Signal(64-LINE_OFF_BITS
)
436 class DTLBUpdate(Elaboratable
):
438 self
.tlbie
= Signal()
439 self
.tlbwe
= Signal()
440 self
.doall
= Signal()
441 self
.tlb_hit
= TLBHit("tlb_hit")
442 self
.tlb_req_index
= Signal(TLB_SET_BITS
)
444 self
.tlb_tag_way
= Signal(TLB_TAG_WAY_BITS
)
445 self
.tlb_pte_way
= Signal(TLB_PTE_WAY_BITS
)
446 self
.repl_way
= Signal(TLB_WAY_BITS
)
447 self
.eatag
= Signal(TLB_EA_TAG_BITS
)
448 self
.pte_data
= Signal(TLB_PTE_BITS
)
450 # read from dtlb array
451 self
.tlb_read
= Signal()
452 self
.tlb_read_index
= Signal(TLB_SET_BITS
)
453 self
.tlb_way
= TLBRecord("o_tlb_way")
455 def elaborate(self
, platform
):
460 # there are 3 parts to this:
461 # QTY TLB_NUM_WAYs TAGs - of width (say) 46 bits of Effective Address
462 # QTY TLB_NUM_WAYs PTEs - of width (say) 64 bits
463 # "Valid" bits, one per "way", of QTY TLB_NUM_WAYs. these cannot
464 # be a Memory because they can all be cleared (tlbie, doall), i mean,
465 # we _could_, in theory, by overriding the Reset Signal of the Memory,
468 dtlb_valid
= TLBValidArray()
469 tlb_req_index
= self
.tlb_req_index
471 print ("TLB_TAG_WAY_BITS", TLB_TAG_WAY_BITS
)
472 print (" TLB_EA_TAG_BITS", TLB_EA_TAG_BITS
)
473 print (" TLB_NUM_WAYS", TLB_NUM_WAYS
)
474 print ("TLB_PTE_WAY_BITS", TLB_PTE_WAY_BITS
)
475 print (" TLB_PTE_BITS", TLB_PTE_BITS
)
476 print (" TLB_NUM_WAYS", TLB_NUM_WAYS
)
478 # TAG and PTE Memory SRAMs. transparent, write-enables are TLB_NUM_WAYS
479 tagway
= Memory(depth
=TLB_SET_SIZE
, width
=TLB_TAG_WAY_BITS
)
480 m
.submodules
.rd_tagway
= rd_tagway
= tagway
.read_port()
481 m
.submodules
.wr_tagway
= wr_tagway
= tagway
.write_port(
482 granularity
=TLB_EA_TAG_BITS
)
484 pteway
= Memory(depth
=TLB_SET_SIZE
, width
=TLB_PTE_WAY_BITS
)
485 m
.submodules
.rd_pteway
= rd_pteway
= pteway
.read_port()
486 m
.submodules
.wr_pteway
= wr_pteway
= pteway
.write_port(
487 granularity
=TLB_PTE_BITS
)
489 # commented out for now, can be put in if Memory.reset can be
490 # used for tlbie&doall to reset the entire Memory to zero in 1 cycle
491 #validm = Memory(depth=TLB_SET_SIZE, width=TLB_NUM_WAYS)
492 #m.submodules.rd_valid = rd_valid = validm.read_port()
493 #m.submodules.wr_valid = wr_valid = validm.write_port(
496 # connect up read and write addresses to Valid/PTE/TAG SRAMs
497 m
.d
.comb
+= rd_pteway
.addr
.eq(self
.tlb_read_index
)
498 m
.d
.comb
+= rd_tagway
.addr
.eq(self
.tlb_read_index
)
499 #m.d.comb += rd_valid.addr.eq(self.tlb_read_index)
500 m
.d
.comb
+= wr_tagway
.addr
.eq(tlb_req_index
)
501 m
.d
.comb
+= wr_pteway
.addr
.eq(tlb_req_index
)
502 #m.d.comb += wr_valid.addr.eq(tlb_req_index)
506 tb_out
= Signal(TLB_TAG_WAY_BITS
) # tlb_way_tags_t
507 db_out
= Signal(TLB_NUM_WAYS
) # tlb_way_valids_t
508 pb_out
= Signal(TLB_PTE_WAY_BITS
) # tlb_way_ptes_t
509 dv
= Signal(TLB_NUM_WAYS
) # tlb_way_valids_t
511 comb
+= dv
.eq(dtlb_valid
[tlb_req_index
])
512 comb
+= db_out
.eq(dv
)
514 with m
.If(self
.tlbie
& self
.doall
):
515 # clear all valid bits at once
516 # XXX hmmm, validm _could_ use Memory reset here...
517 for i
in range(TLB_SET_SIZE
):
518 sync
+= dtlb_valid
[i
].eq(0)
519 with m
.Elif(self
.tlbie
):
520 # invalidate just the hit_way
521 with m
.If(self
.tlb_hit
.valid
):
522 comb
+= db_out
.bit_select(self
.tlb_hit
.way
, 1).eq(0)
523 comb
+= v_updated
.eq(1)
524 with m
.Elif(self
.tlbwe
):
525 # write to the requested tag and PTE
526 comb
+= write_tlb_tag(self
.repl_way
, tb_out
, self
.eatag
)
527 comb
+= write_tlb_pte(self
.repl_way
, pb_out
, self
.pte_data
)
529 comb
+= db_out
.bit_select(self
.repl_way
, 1).eq(1)
531 comb
+= updated
.eq(1)
532 comb
+= v_updated
.eq(1)
534 # above, sometimes valid is requested to be updated but data not
535 # therefore split them out, here. note the granularity thing matches
536 # with the shift-up of the eatag/pte_data into the correct TLB way.
537 # thus is it not necessary to write the entire lot, just the portion
538 # being altered: hence writing the *old* copy of the row is not needed
539 with m
.If(updated
): # PTE and TAG to be written
540 comb
+= wr_pteway
.data
.eq(pb_out
)
541 comb
+= wr_pteway
.en
.eq(1<<self
.repl_way
)
542 comb
+= wr_tagway
.data
.eq(tb_out
)
543 comb
+= wr_tagway
.en
.eq(1<<self
.repl_way
)
544 with m
.If(v_updated
): # Valid to be written
545 sync
+= dtlb_valid
[tlb_req_index
].eq(db_out
)
546 #comb += wr_valid.data.eq(db_out)
547 #comb += wr_valid.en.eq(1<<self.repl_way)
549 # select one TLB way, use a register here
550 r_tlb_way
= TLBRecord("r_tlb_way")
552 sync
+= r_delay
.eq(self
.tlb_read
)
553 with m
.If(self
.tlb_read
):
554 sync
+= self
.tlb_way
.valid
.eq(dtlb_valid
[self
.tlb_read_index
])
556 # on one clock delay, output the contents of the read port(s)
557 # comb += self.tlb_way.valid.eq(rd_valid.data)
558 comb
+= self
.tlb_way
.tag
.eq(rd_tagway
.data
)
559 comb
+= self
.tlb_way
.pte
.eq(rd_pteway
.data
)
560 # and also capture the (delayed) output...
561 #sync += r_tlb_way.valid.eq(rd_valid.data)
562 sync
+= r_tlb_way
.tag
.eq(rd_tagway
.data
)
563 sync
+= r_tlb_way
.pte
.eq(rd_pteway
.data
)
565 # ... so that the register can output it when no read is requested
566 # it's rather overkill but better to be safe than sorry
567 comb
+= self
.tlb_way
.tag
.eq(r_tlb_way
.tag
)
568 comb
+= self
.tlb_way
.pte
.eq(r_tlb_way
.pte
)
569 #comb += self.tlb_way.eq(r_tlb_way)
574 class DCachePendingHit(Elaboratable
):
576 def __init__(self
, tlb_way
,
577 cache_i_validdx
, cache_tag_set
,
582 self
.virt_mode
= Signal()
583 self
.is_hit
= Signal()
584 self
.tlb_hit
= TLBHit("tlb_hit")
585 self
.hit_way
= Signal(WAY_BITS
)
586 self
.rel_match
= Signal()
587 self
.req_index
= Signal(INDEX_BITS
)
588 self
.reload_tag
= Signal(TAG_BITS
)
590 self
.tlb_way
= tlb_way
591 self
.cache_i_validdx
= cache_i_validdx
592 self
.cache_tag_set
= cache_tag_set
593 self
.req_addr
= req_addr
594 self
.hit_set
= hit_set
596 def elaborate(self
, platform
):
602 virt_mode
= self
.virt_mode
604 tlb_way
= self
.tlb_way
605 cache_i_validdx
= self
.cache_i_validdx
606 cache_tag_set
= self
.cache_tag_set
607 req_addr
= self
.req_addr
608 tlb_hit
= self
.tlb_hit
609 hit_set
= self
.hit_set
610 hit_way
= self
.hit_way
611 rel_match
= self
.rel_match
612 req_index
= self
.req_index
613 reload_tag
= self
.reload_tag
615 rel_matches
= Array(Signal(name
="rel_matches_%d" % i
) \
616 for i
in range(TLB_NUM_WAYS
))
617 hit_way_set
= HitWaySet()
619 # Test if pending request is a hit on any way
620 # In order to make timing in virtual mode,
621 # when we are using the TLB, we compare each
622 # way with each of the real addresses from each way of
623 # the TLB, and then decide later which match to use.
625 with m
.If(virt_mode
):
626 for j
in range(TLB_NUM_WAYS
): # tlb_num_way_t
627 s_tag
= Signal(TAG_BITS
, name
="s_tag%d" % j
)
629 s_pte
= Signal(TLB_PTE_BITS
)
630 s_ra
= Signal(REAL_ADDR_BITS
)
631 comb
+= s_pte
.eq(read_tlb_pte(j
, tlb_way
.pte
))
632 comb
+= s_ra
.eq(Cat(req_addr
[0:TLB_LG_PGSZ
],
633 s_pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
634 comb
+= s_tag
.eq(get_tag(s_ra
))
636 for i
in range(NUM_WAYS
): # way_t
637 is_tag_hit
= Signal(name
="is_tag_hit_%d_%d" % (j
, i
))
638 comb
+= is_tag_hit
.eq(go
& cache_i_validdx
[i
] &
639 (read_tag(i
, cache_tag_set
) == s_tag
)
640 & (tlb_way
.valid
[j
]))
641 with m
.If(is_tag_hit
):
642 comb
+= hit_way_set
[j
].eq(i
)
644 comb
+= hit_set
[j
].eq(s_hit
)
645 with m
.If(s_tag
== reload_tag
):
646 comb
+= rel_matches
[j
].eq(1)
647 with m
.If(tlb_hit
.valid
):
648 comb
+= is_hit
.eq(hit_set
[tlb_hit
.way
])
649 comb
+= hit_way
.eq(hit_way_set
[tlb_hit
.way
])
650 comb
+= rel_match
.eq(rel_matches
[tlb_hit
.way
])
652 s_tag
= Signal(TAG_BITS
)
653 comb
+= s_tag
.eq(get_tag(req_addr
))
654 for i
in range(NUM_WAYS
): # way_t
655 is_tag_hit
= Signal(name
="is_tag_hit_%d" % i
)
656 comb
+= is_tag_hit
.eq(go
& cache_i_validdx
[i
] &
657 (read_tag(i
, cache_tag_set
) == s_tag
))
658 with m
.If(is_tag_hit
):
659 comb
+= hit_way
.eq(i
)
661 with m
.If(s_tag
== reload_tag
):
662 comb
+= rel_match
.eq(1)
667 class DCache(Elaboratable
):
668 """Set associative dcache write-through
670 TODO (in no specific order):
671 * See list in icache.vhdl
672 * Complete load misses on the cycle when WB data comes instead of
673 at the end of line (this requires dealing with requests coming in
677 self
.d_in
= LoadStore1ToDCacheType("d_in")
678 self
.d_out
= DCacheToLoadStore1Type("d_out")
680 self
.m_in
= MMUToDCacheType("m_in")
681 self
.m_out
= DCacheToMMUType("m_out")
683 self
.stall_out
= Signal()
685 # standard naming (wired to non-standard for compatibility)
686 self
.bus
= Interface(addr_width
=32,
693 self
.log_out
= Signal(20)
695 def stage_0(self
, m
, r0
, r1
, r0_full
):
696 """Latch the request in r0.req as long as we're not stalling
700 d_in
, d_out
, m_in
= self
.d_in
, self
.d_out
, self
.m_in
702 r
= RegStage0("stage0")
704 # TODO, this goes in unit tests and formal proofs
705 with m
.If(d_in
.valid
& m_in
.valid
):
706 sync
+= Display("request collision loadstore vs MMU")
708 with m
.If(m_in
.valid
):
709 comb
+= r
.req
.valid
.eq(1)
710 comb
+= r
.req
.load
.eq(~
(m_in
.tlbie | m_in
.tlbld
))# no invalidate
711 comb
+= r
.req
.dcbz
.eq(0)
712 comb
+= r
.req
.nc
.eq(0)
713 comb
+= r
.req
.reserve
.eq(0)
714 comb
+= r
.req
.virt_mode
.eq(0)
715 comb
+= r
.req
.priv_mode
.eq(1)
716 comb
+= r
.req
.addr
.eq(m_in
.addr
)
717 comb
+= r
.req
.data
.eq(m_in
.pte
)
718 comb
+= r
.req
.byte_sel
.eq(~
0) # Const -1 sets all to 0b111....
719 comb
+= r
.tlbie
.eq(m_in
.tlbie
)
720 comb
+= r
.doall
.eq(m_in
.doall
)
721 comb
+= r
.tlbld
.eq(m_in
.tlbld
)
722 comb
+= r
.mmu_req
.eq(1)
723 m
.d
.sync
+= Display(" DCACHE req mmu addr %x pte %x ld %d",
724 m_in
.addr
, m_in
.pte
, r
.req
.load
)
727 comb
+= r
.req
.eq(d_in
)
728 comb
+= r
.req
.data
.eq(0)
729 comb
+= r
.tlbie
.eq(0)
730 comb
+= r
.doall
.eq(0)
731 comb
+= r
.tlbld
.eq(0)
732 comb
+= r
.mmu_req
.eq(0)
733 with m
.If((~r1
.full
& ~d_in
.hold
) | ~r0_full
):
735 sync
+= r0_full
.eq(r
.req
.valid
)
736 # Sample data the cycle after a request comes in from loadstore1.
737 # If another request has come in already then the data will get
738 # put directly into req.data below.
739 with m
.If(r0
.req
.valid
& ~r
.req
.valid
& ~r0
.d_valid
&
741 sync
+= r0
.req
.data
.eq(d_in
.data
)
742 sync
+= r0
.d_valid
.eq(1)
743 with m
.If(d_in
.valid
):
744 m
.d
.sync
+= Display(" DCACHE req cache "
745 "virt %d addr %x data %x ld %d",
746 r
.req
.virt_mode
, r
.req
.addr
,
747 r
.req
.data
, r
.req
.load
)
749 def tlb_read(self
, m
, r0_stall
, tlb_way
):
751 Operates in the second cycle on the request latched in r0.req.
752 TLB updates write the entry at the end of the second cycle.
756 m_in
, d_in
= self
.m_in
, self
.d_in
758 addrbits
= Signal(TLB_SET_BITS
)
761 amax
= TLB_LG_PGSZ
+ TLB_SET_BITS
763 with m
.If(m_in
.valid
):
764 comb
+= addrbits
.eq(m_in
.addr
[amin
: amax
])
766 comb
+= addrbits
.eq(d_in
.addr
[amin
: amax
])
768 # If we have any op and the previous op isn't finished,
769 # then keep the same output for next cycle.
771 comb
+= d
.tlb_read_index
.eq(addrbits
)
772 comb
+= d
.tlb_read
.eq(~r0_stall
)
773 comb
+= tlb_way
.eq(d
.tlb_way
)
775 def maybe_tlb_plrus(self
, m
, r1
, tlb_plru_victim
, tlb_req_index
):
776 """Generate TLB PLRUs
781 if TLB_NUM_WAYS
== 0:
784 # Binary-to-Unary one-hot, enabled by tlb_hit valid
785 tlb_plrus
= PLRUs(TLB_SET_SIZE
, TLB_WAY_BITS
)
786 m
.submodules
.tlb_plrus
= tlb_plrus
787 comb
+= tlb_plrus
.way
.eq(r1
.tlb_hit
.way
)
788 comb
+= tlb_plrus
.valid
.eq(r1
.tlb_hit
.valid
)
789 comb
+= tlb_plrus
.index
.eq(r1
.tlb_hit_index
)
790 comb
+= tlb_plrus
.isel
.eq(tlb_req_index
) # select victim
791 comb
+= tlb_plru_victim
.eq(tlb_plrus
.o_index
) # selected victim
793 def tlb_search(self
, m
, tlb_req_index
, r0
, r0_valid
,
795 pte
, tlb_hit
, valid_ra
, perm_attr
, ra
):
799 hitway
= Signal(TLB_WAY_BITS
)
801 eatag
= Signal(TLB_EA_TAG_BITS
)
803 TLB_LG_END
= TLB_LG_PGSZ
+ TLB_SET_BITS
804 comb
+= tlb_req_index
.eq(r0
.req
.addr
[TLB_LG_PGSZ
: TLB_LG_END
])
805 comb
+= eatag
.eq(r0
.req
.addr
[TLB_LG_END
: 64 ])
807 for i
in range(TLB_NUM_WAYS
):
808 is_tag_hit
= Signal(name
="is_tag_hit%d" % i
)
809 tlb_tag
= Signal(TLB_EA_TAG_BITS
, name
="tlb_tag%d" % i
)
810 comb
+= tlb_tag
.eq(read_tlb_tag(i
, tlb_way
.tag
))
811 comb
+= is_tag_hit
.eq((tlb_way
.valid
[i
]) & (tlb_tag
== eatag
))
812 with m
.If(is_tag_hit
):
816 comb
+= tlb_hit
.valid
.eq(hit
& r0_valid
)
817 comb
+= tlb_hit
.way
.eq(hitway
)
819 with m
.If(tlb_hit
.valid
):
820 comb
+= pte
.eq(read_tlb_pte(hitway
, tlb_way
.pte
))
821 comb
+= valid_ra
.eq(tlb_hit
.valid | ~r0
.req
.virt_mode
)
823 with m
.If(r0
.req
.virt_mode
):
824 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
825 r0
.req
.addr
[ROW_OFF_BITS
:TLB_LG_PGSZ
],
826 pte
[TLB_LG_PGSZ
:REAL_ADDR_BITS
]))
827 comb
+= perm_attr
.reference
.eq(pte
[8])
828 comb
+= perm_attr
.changed
.eq(pte
[7])
829 comb
+= perm_attr
.nocache
.eq(pte
[5])
830 comb
+= perm_attr
.priv
.eq(pte
[3])
831 comb
+= perm_attr
.rd_perm
.eq(pte
[2])
832 comb
+= perm_attr
.wr_perm
.eq(pte
[1])
834 comb
+= ra
.eq(Cat(Const(0, ROW_OFF_BITS
),
835 r0
.req
.addr
[ROW_OFF_BITS
:REAL_ADDR_BITS
]))
836 comb
+= perm_attr
.reference
.eq(1)
837 comb
+= perm_attr
.changed
.eq(1)
838 comb
+= perm_attr
.nocache
.eq(0)
839 comb
+= perm_attr
.priv
.eq(1)
840 comb
+= perm_attr
.rd_perm
.eq(1)
841 comb
+= perm_attr
.wr_perm
.eq(1)
844 m
.d
.sync
+= Display("DCACHE virt mode %d hit %d ra %x pte %x",
845 r0
.req
.virt_mode
, tlb_hit
.valid
, ra
, pte
)
846 m
.d
.sync
+= Display(" perm ref=%d", perm_attr
.reference
)
847 m
.d
.sync
+= Display(" perm chg=%d", perm_attr
.changed
)
848 m
.d
.sync
+= Display(" perm noc=%d", perm_attr
.nocache
)
849 m
.d
.sync
+= Display(" perm prv=%d", perm_attr
.priv
)
850 m
.d
.sync
+= Display(" perm rdp=%d", perm_attr
.rd_perm
)
851 m
.d
.sync
+= Display(" perm wrp=%d", perm_attr
.wr_perm
)
853 def tlb_update(self
, m
, r0_valid
, r0
, tlb_req_index
,
854 tlb_hit
, tlb_plru_victim
, tlb_way
):
862 comb
+= tlbie
.eq(r0_valid
& r0
.tlbie
)
863 comb
+= tlbwe
.eq(r0_valid
& r0
.tlbld
)
867 comb
+= d
.tlbie
.eq(tlbie
)
868 comb
+= d
.tlbwe
.eq(tlbwe
)
869 comb
+= d
.doall
.eq(r0
.doall
)
870 comb
+= d
.tlb_hit
.eq(tlb_hit
)
871 comb
+= d
.tlb_tag_way
.eq(tlb_way
.tag
)
872 comb
+= d
.tlb_pte_way
.eq(tlb_way
.pte
)
873 comb
+= d
.tlb_req_index
.eq(tlb_req_index
)
875 with m
.If(tlb_hit
.valid
):
876 comb
+= d
.repl_way
.eq(tlb_hit
.way
)
878 comb
+= d
.repl_way
.eq(tlb_plru_victim
)
879 comb
+= d
.eatag
.eq(r0
.req
.addr
[TLB_LG_PGSZ
+ TLB_SET_BITS
:64])
880 comb
+= d
.pte_data
.eq(r0
.req
.data
)
882 def maybe_plrus(self
, m
, r1
, plru_victim
):
888 if TLB_NUM_WAYS
== 0:
891 m
.submodules
.plrus
= plrus
= PLRUs(NUM_LINES
, WAY_BITS
)
892 comb
+= plrus
.way
.eq(r1
.hit_way
)
893 comb
+= plrus
.valid
.eq(r1
.cache_hit
)
894 comb
+= plrus
.index
.eq(r1
.hit_index
)
895 comb
+= plrus
.isel
.eq(r1
.store_index
) # select victim
896 comb
+= plru_victim
.eq(plrus
.o_index
) # selected victim
898 def cache_tag_read(self
, m
, r0_stall
, req_index
, cache_tag_set
, cache_tags
):
899 """Cache tag RAM read port
903 m_in
, d_in
= self
.m_in
, self
.d_in
905 index
= Signal(INDEX_BITS
)
908 comb
+= index
.eq(req_index
)
909 with m
.Elif(m_in
.valid
):
910 comb
+= index
.eq(get_index(m_in
.addr
))
912 comb
+= index
.eq(get_index(d_in
.addr
))
913 sync
+= cache_tag_set
.eq(cache_tags
[index
].tag
)
915 def dcache_request(self
, m
, r0
, ra
, req_index
, req_row
, req_tag
,
916 r0_valid
, r1
, cache_tags
, replace_way
,
917 use_forward1_next
, use_forward2_next
,
918 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
919 valid_ra
, perm_ok
, access_ok
, req_op
, req_go
,
920 tlb_hit
, tlb_way
, cache_tag_set
,
921 cancel_store
, req_same_tag
, r0_stall
, early_req_row
):
922 """Cache request parsing and hit detection
926 m_in
, d_in
= self
.m_in
, self
.d_in
929 hit_way
= Signal(WAY_BITS
)
934 hit_set
= Array(Signal(name
="hit_set_%d" % i
) \
935 for i
in range(TLB_NUM_WAYS
))
936 cache_i_validdx
= Signal(NUM_WAYS
)
938 # Extract line, row and tag from request
939 comb
+= req_index
.eq(get_index(r0
.req
.addr
))
940 comb
+= req_row
.eq(get_row(r0
.req
.addr
))
941 comb
+= req_tag
.eq(get_tag(ra
))
943 if False: # display on comb is a bit... busy.
944 comb
+= Display("dcache_req addr:%x ra: %x idx: %x tag: %x row: %x",
945 r0
.req
.addr
, ra
, req_index
, req_tag
, req_row
)
947 comb
+= go
.eq(r0_valid
& ~
(r0
.tlbie | r0
.tlbld
) & ~r1
.ls_error
)
948 comb
+= cache_i_validdx
.eq(cache_tags
[req_index
].valid
)
950 m
.submodules
.dcache_pend
= dc
= DCachePendingHit(tlb_way
,
951 cache_i_validdx
, cache_tag_set
,
954 comb
+= dc
.tlb_hit
.eq(tlb_hit
)
955 comb
+= dc
.reload_tag
.eq(r1
.reload_tag
)
956 comb
+= dc
.virt_mode
.eq(r0
.req
.virt_mode
)
958 comb
+= dc
.req_index
.eq(req_index
)
960 comb
+= is_hit
.eq(dc
.is_hit
)
961 comb
+= hit_way
.eq(dc
.hit_way
)
962 comb
+= req_same_tag
.eq(dc
.rel_match
)
964 # See if the request matches the line currently being reloaded
965 with m
.If((r1
.state
== State
.RELOAD_WAIT_ACK
) &
966 (req_index
== r1
.store_index
) & req_same_tag
):
967 # For a store, consider this a hit even if the row isn't
968 # valid since it will be by the time we perform the store.
969 # For a load, check the appropriate row valid bit.
970 rrow
= Signal(ROW_LINE_BITS
)
971 comb
+= rrow
.eq(req_row
)
972 valid
= r1
.rows_valid
[rrow
]
973 comb
+= is_hit
.eq((~r0
.req
.load
) | valid
)
974 comb
+= hit_way
.eq(replace_way
)
976 # Whether to use forwarded data for a load or not
977 with m
.If((get_row(r1
.req
.real_addr
) == req_row
) &
978 (r1
.req
.hit_way
== hit_way
)):
979 # Only need to consider r1.write_bram here, since if we
980 # are writing refill data here, then we don't have a
981 # cache hit this cycle on the line being refilled.
982 # (There is the possibility that the load following the
983 # load miss that started the refill could be to the old
984 # contents of the victim line, since it is a couple of
985 # cycles after the refill starts before we see the updated
986 # cache tag. In that case we don't use the bypass.)
987 comb
+= use_forward1_next
.eq(r1
.write_bram
)
988 with m
.If((r1
.forward_row1
== req_row
) & (r1
.forward_way1
== hit_way
)):
989 comb
+= use_forward2_next
.eq(r1
.forward_valid1
)
991 # The way that matched on a hit
992 comb
+= req_hit_way
.eq(hit_way
)
994 # The way to replace on a miss
995 with m
.If(r1
.write_tag
):
996 comb
+= replace_way
.eq(plru_victim
)
998 comb
+= replace_way
.eq(r1
.store_way
)
1000 # work out whether we have permission for this access
1001 # NB we don't yet implement AMR, thus no KUAP
1002 comb
+= rc_ok
.eq(perm_attr
.reference
1003 & (r0
.req
.load | perm_attr
.changed
))
1004 comb
+= perm_ok
.eq((r0
.req
.priv_mode |
(~perm_attr
.priv
)) &
1005 (perm_attr
.wr_perm |
1006 (r0
.req
.load
& perm_attr
.rd_perm
)))
1007 comb
+= access_ok
.eq(valid_ra
& perm_ok
& rc_ok
)
1009 # Combine the request and cache hit status to decide what
1010 # operation needs to be done
1011 comb
+= nc
.eq(r0
.req
.nc | perm_attr
.nocache
)
1012 comb
+= op
.eq(Op
.OP_NONE
)
1014 with m
.If(~access_ok
):
1015 m
.d
.sync
+= Display("DCACHE access fail valid_ra=%d p=%d rc=%d",
1016 valid_ra
, perm_ok
, rc_ok
)
1017 comb
+= op
.eq(Op
.OP_BAD
)
1018 with m
.Elif(cancel_store
):
1019 m
.d
.sync
+= Display("DCACHE cancel store")
1020 comb
+= op
.eq(Op
.OP_STCX_FAIL
)
1022 m
.d
.sync
+= Display("DCACHE valid_ra=%d nc=%d ld=%d",
1023 valid_ra
, nc
, r0
.req
.load
)
1024 comb
+= opsel
.eq(Cat(is_hit
, nc
, r0
.req
.load
))
1025 with m
.Switch(opsel
):
1026 with m
.Case(0b101): comb
+= op
.eq(Op
.OP_LOAD_HIT
)
1027 with m
.Case(0b100): comb
+= op
.eq(Op
.OP_LOAD_MISS
)
1028 with m
.Case(0b110): comb
+= op
.eq(Op
.OP_LOAD_NC
)
1029 with m
.Case(0b001): comb
+= op
.eq(Op
.OP_STORE_HIT
)
1030 with m
.Case(0b000): comb
+= op
.eq(Op
.OP_STORE_MISS
)
1031 with m
.Case(0b010): comb
+= op
.eq(Op
.OP_STORE_MISS
)
1032 with m
.Case(0b011): comb
+= op
.eq(Op
.OP_BAD
)
1033 with m
.Case(0b111): comb
+= op
.eq(Op
.OP_BAD
)
1034 comb
+= req_op
.eq(op
)
1035 comb
+= req_go
.eq(go
)
1037 # Version of the row number that is valid one cycle earlier
1038 # in the cases where we need to read the cache data BRAM.
1039 # If we're stalling then we need to keep reading the last
1041 with m
.If(~r0_stall
):
1042 with m
.If(m_in
.valid
):
1043 comb
+= early_req_row
.eq(get_row(m_in
.addr
))
1045 comb
+= early_req_row
.eq(get_row(d_in
.addr
))
1047 comb
+= early_req_row
.eq(req_row
)
1049 def reservation_comb(self
, m
, cancel_store
, set_rsrv
, clear_rsrv
,
1050 r0_valid
, r0
, reservation
):
1051 """Handle load-with-reservation and store-conditional instructions
1055 with m
.If(r0_valid
& r0
.req
.reserve
):
1056 # XXX generate alignment interrupt if address
1057 # is not aligned XXX or if r0.req.nc = '1'
1058 with m
.If(r0
.req
.load
):
1059 comb
+= set_rsrv
.eq(r0
.req
.atomic_last
) # load with reservation
1061 comb
+= clear_rsrv
.eq(r0
.req
.atomic_last
) # store conditional
1062 with m
.If((~reservation
.valid
) |
1063 (r0
.req
.addr
[LINE_OFF_BITS
:64] != reservation
.addr
)):
1064 comb
+= cancel_store
.eq(1)
1066 def reservation_reg(self
, m
, r0_valid
, access_ok
, set_rsrv
, clear_rsrv
,
1071 with m
.If(r0_valid
& access_ok
):
1072 with m
.If(clear_rsrv
):
1073 sync
+= reservation
.valid
.eq(0)
1074 with m
.Elif(set_rsrv
):
1075 sync
+= reservation
.valid
.eq(1)
1076 sync
+= reservation
.addr
.eq(r0
.req
.addr
[LINE_OFF_BITS
:64])
1078 def writeback_control(self
, m
, r1
, cache_out_row
):
1079 """Return data for loads & completion control logic
1083 d_out
, m_out
= self
.d_out
, self
.m_out
1085 data_out
= Signal(64)
1086 data_fwd
= Signal(64)
1088 # Use the bypass if are reading the row that was
1089 # written 1 or 2 cycles ago, including for the
1090 # slow_valid = 1 case (i.e. completing a load
1091 # miss or a non-cacheable load).
1092 with m
.If(r1
.use_forward1
):
1093 comb
+= data_fwd
.eq(r1
.forward_data1
)
1095 comb
+= data_fwd
.eq(r1
.forward_data2
)
1097 comb
+= data_out
.eq(cache_out_row
)
1100 with m
.If(r1
.forward_sel
[i
]):
1101 dsel
= data_fwd
.word_select(i
, 8)
1102 comb
+= data_out
.word_select(i
, 8).eq(dsel
)
1104 # DCache output to LoadStore
1105 comb
+= d_out
.valid
.eq(r1
.ls_valid
)
1106 comb
+= d_out
.data
.eq(data_out
)
1107 comb
+= d_out
.store_done
.eq(~r1
.stcx_fail
)
1108 comb
+= d_out
.error
.eq(r1
.ls_error
)
1109 comb
+= d_out
.cache_paradox
.eq(r1
.cache_paradox
)
1112 comb
+= m_out
.done
.eq(r1
.mmu_done
)
1113 comb
+= m_out
.err
.eq(r1
.mmu_error
)
1114 comb
+= m_out
.data
.eq(data_out
)
1116 # We have a valid load or store hit or we just completed
1117 # a slow op such as a load miss, a NC load or a store
1119 # Note: the load hit is delayed by one cycle. However it
1120 # can still not collide with r.slow_valid (well unless I
1121 # miscalculated) because slow_valid can only be set on a
1122 # subsequent request and not on its first cycle (the state
1123 # machine must have advanced), which makes slow_valid
1124 # at least 2 cycles from the previous hit_load_valid.
1126 # Sanity: Only one of these must be set in any given cycle
1128 if False: # TODO: need Display to get this to work
1129 assert (r1
.slow_valid
& r1
.stcx_fail
) != 1, \
1130 "unexpected slow_valid collision with stcx_fail"
1132 assert ((r1
.slow_valid | r1
.stcx_fail
) | r1
.hit_load_valid
) != 1, \
1133 "unexpected hit_load_delayed collision with slow_valid"
1135 with m
.If(~r1
.mmu_req
):
1136 # Request came from loadstore1...
1137 # Load hit case is the standard path
1138 with m
.If(r1
.hit_load_valid
):
1139 sync
+= Display("completing load hit data=%x", data_out
)
1141 # error cases complete without stalling
1142 with m
.If(r1
.ls_error
):
1144 sync
+= Display("completing dcbz with error")
1146 sync
+= Display("completing ld/st with error")
1148 # Slow ops (load miss, NC, stores)
1149 with m
.If(r1
.slow_valid
):
1150 sync
+= Display("completing store or load miss adr=%x data=%x",
1151 r1
.req
.real_addr
, data_out
)
1154 # Request came from MMU
1155 with m
.If(r1
.hit_load_valid
):
1156 sync
+= Display("completing load hit to MMU, data=%x",
1158 # error cases complete without stalling
1159 with m
.If(r1
.mmu_error
):
1160 sync
+= Display("combpleting MMU ld with error")
1162 # Slow ops (i.e. load miss)
1163 with m
.If(r1
.slow_valid
):
1164 sync
+= Display("completing MMU load miss, adr=%x data=%x",
1165 r1
.req
.real_addr
, m_out
.data
)
1167 def rams(self
, m
, r1
, early_req_row
, cache_out_row
, replace_way
):
1169 Generate a cache RAM for each way. This handles the normal
1170 reads, writes from reloads and the special store-hit update
1173 Note: the BRAMs have an extra read buffer, meaning the output
1174 is pipelined an extra cycle. This differs from the
1175 icache. The writeback logic needs to take that into
1176 account by using 1-cycle delayed signals for load hits.
1181 # a Binary-to-Unary one-hots here. replace-way one-hot is gated
1182 # (enabled) by bus.ack, not-write-bram, and state RELOAD_WAIT_ACK
1183 m
.submodules
.rams_replace_way_e
= rwe
= Decoder(NUM_WAYS
)
1184 comb
+= rwe
.n
.eq(~
((r1
.state
== State
.RELOAD_WAIT_ACK
) & bus
.ack
&
1186 comb
+= rwe
.i
.eq(replace_way
)
1188 m
.submodules
.rams_hit_way_e
= hwe
= Decoder(NUM_WAYS
)
1189 comb
+= hwe
.i
.eq(r1
.hit_way
)
1191 # this one is gated with write_bram, and replace_way_e can never be
1192 # set at the same time. that means that do_write can OR the outputs
1193 m
.submodules
.rams_hit_req_way_e
= hre
= Decoder(NUM_WAYS
)
1194 comb
+= hre
.n
.eq(~r1
.write_bram
) # Decoder.n is inverted
1195 comb
+= hre
.i
.eq(r1
.req
.hit_way
)
1199 wr_addr
= Signal(ROW_BITS
)
1200 wr_data
= Signal(WB_DATA_BITS
)
1201 wr_sel
= Signal(ROW_SIZE
)
1202 rd_addr
= Signal(ROW_BITS
)
1204 comb
+= do_read
.eq(1) # always enable
1205 comb
+= rd_addr
.eq(early_req_row
)
1209 # Defaults to wishbone read responses (cache refill)
1211 # For timing, the mux on wr_data/sel/addr is not
1212 # dependent on anything other than the current state.
1214 with m
.If(r1
.write_bram
):
1215 # Write store data to BRAM. This happens one
1216 # cycle after the store is in r0.
1217 comb
+= wr_data
.eq(r1
.req
.data
)
1218 comb
+= wr_sel
.eq(r1
.req
.byte_sel
)
1219 comb
+= wr_addr
.eq(get_row(r1
.req
.real_addr
))
1222 # Otherwise, we might be doing a reload or a DCBZ
1224 comb
+= wr_data
.eq(0)
1226 comb
+= wr_data
.eq(bus
.dat_r
)
1227 comb
+= wr_addr
.eq(r1
.store_row
)
1228 comb
+= wr_sel
.eq(~
0) # all 1s
1231 for i
in range(NUM_WAYS
):
1232 do_write
= Signal(name
="do_wr%d" % i
)
1233 wr_sel_m
= Signal(ROW_SIZE
, name
="wr_sel_m_%d" % i
)
1234 d_out
= Signal(WB_DATA_BITS
, name
="dout_%d" % i
) # cache_row_t
1236 way
= CacheRam(ROW_BITS
, WB_DATA_BITS
, ADD_BUF
=True, ram_num
=i
)
1237 setattr(m
.submodules
, "cacheram_%d" % i
, way
)
1239 comb
+= way
.rd_en
.eq(do_read
)
1240 comb
+= way
.rd_addr
.eq(rd_addr
)
1241 comb
+= d_out
.eq(way
.rd_data_o
)
1242 comb
+= way
.wr_sel
.eq(wr_sel_m
)
1243 comb
+= way
.wr_addr
.eq(wr_addr
)
1244 comb
+= way
.wr_data
.eq(wr_data
)
1247 with m
.If(hwe
.o
[i
]):
1248 comb
+= cache_out_row
.eq(d_out
)
1250 # these are mutually-exclusive via their Decoder-enablers
1251 # (note: Decoder-enable is inverted)
1252 comb
+= do_write
.eq(hre
.o
[i
] | rwe
.o
[i
])
1254 # Mask write selects with do_write since BRAM
1255 # doesn't have a global write-enable
1256 with m
.If(do_write
):
1257 comb
+= wr_sel_m
.eq(wr_sel
)
1259 # Cache hit synchronous machine for the easy case.
1260 # This handles load hits.
1261 # It also handles error cases (TLB miss, cache paradox)
1262 def dcache_fast_hit(self
, m
, req_op
, r0_valid
, r0
, r1
,
1263 req_hit_way
, req_index
, req_tag
, access_ok
,
1264 tlb_hit
, tlb_req_index
):
1268 with m
.If(req_op
!= Op
.OP_NONE
):
1269 sync
+= Display("op:%d addr:%x nc: %d idx: %x tag: %x way: %x",
1270 req_op
, r0
.req
.addr
, r0
.req
.nc
,
1271 req_index
, req_tag
, req_hit_way
)
1273 with m
.If(r0_valid
):
1274 sync
+= r1
.mmu_req
.eq(r0
.mmu_req
)
1276 # Fast path for load/store hits.
1277 # Set signals for the writeback controls.
1278 sync
+= r1
.hit_way
.eq(req_hit_way
)
1279 sync
+= r1
.hit_index
.eq(req_index
)
1281 sync
+= r1
.hit_load_valid
.eq(req_op
== Op
.OP_LOAD_HIT
)
1282 sync
+= r1
.cache_hit
.eq((req_op
== Op
.OP_LOAD_HIT
) |
1283 (req_op
== Op
.OP_STORE_HIT
))
1285 with m
.If(req_op
== Op
.OP_BAD
):
1286 sync
+= Display("Signalling ld/st error "
1287 "ls_error=%i mmu_error=%i cache_paradox=%i",
1288 ~r0
.mmu_req
,r0
.mmu_req
,access_ok
)
1289 sync
+= r1
.ls_error
.eq(~r0
.mmu_req
)
1290 sync
+= r1
.mmu_error
.eq(r0
.mmu_req
)
1291 sync
+= r1
.cache_paradox
.eq(access_ok
)
1293 sync
+= r1
.ls_error
.eq(0)
1294 sync
+= r1
.mmu_error
.eq(0)
1295 sync
+= r1
.cache_paradox
.eq(0)
1297 sync
+= r1
.stcx_fail
.eq(req_op
== Op
.OP_STCX_FAIL
)
1299 # Record TLB hit information for updating TLB PLRU
1300 sync
+= r1
.tlb_hit
.eq(tlb_hit
)
1301 sync
+= r1
.tlb_hit_index
.eq(tlb_req_index
)
1303 # Memory accesses are handled by this state machine:
1305 # * Cache load miss/reload (in conjunction with "rams")
1306 # * Load hits for non-cachable forms
1307 # * Stores (the collision case is handled in "rams")
1309 # All wishbone requests generation is done here.
1310 # This machine operates at stage 1.
1311 def dcache_slow(self
, m
, r1
, use_forward1_next
, use_forward2_next
,
1313 req_hit_way
, req_same_tag
,
1314 r0_valid
, req_op
, cache_tags
, req_go
, ra
):
1321 req
= MemAccessRequest("mreq_ds")
1323 req_row
= Signal(ROW_BITS
)
1324 req_idx
= Signal(INDEX_BITS
)
1325 req_tag
= Signal(TAG_BITS
)
1326 comb
+= req_idx
.eq(get_index(req
.real_addr
))
1327 comb
+= req_row
.eq(get_row(req
.real_addr
))
1328 comb
+= req_tag
.eq(get_tag(req
.real_addr
))
1330 sync
+= r1
.use_forward1
.eq(use_forward1_next
)
1331 sync
+= r1
.forward_sel
.eq(0)
1333 with m
.If(use_forward1_next
):
1334 sync
+= r1
.forward_sel
.eq(r1
.req
.byte_sel
)
1335 with m
.Elif(use_forward2_next
):
1336 sync
+= r1
.forward_sel
.eq(r1
.forward_sel1
)
1338 sync
+= r1
.forward_data2
.eq(r1
.forward_data1
)
1339 with m
.If(r1
.write_bram
):
1340 sync
+= r1
.forward_data1
.eq(r1
.req
.data
)
1341 sync
+= r1
.forward_sel1
.eq(r1
.req
.byte_sel
)
1342 sync
+= r1
.forward_way1
.eq(r1
.req
.hit_way
)
1343 sync
+= r1
.forward_row1
.eq(get_row(r1
.req
.real_addr
))
1344 sync
+= r1
.forward_valid1
.eq(1)
1347 sync
+= r1
.forward_data1
.eq(0)
1349 sync
+= r1
.forward_data1
.eq(bus
.dat_r
)
1350 sync
+= r1
.forward_sel1
.eq(~
0) # all 1s
1351 sync
+= r1
.forward_way1
.eq(replace_way
)
1352 sync
+= r1
.forward_row1
.eq(r1
.store_row
)
1353 sync
+= r1
.forward_valid1
.eq(0)
1355 # One cycle pulses reset
1356 sync
+= r1
.slow_valid
.eq(0)
1357 sync
+= r1
.write_bram
.eq(0)
1358 sync
+= r1
.inc_acks
.eq(0)
1359 sync
+= r1
.dec_acks
.eq(0)
1361 sync
+= r1
.ls_valid
.eq(0)
1362 # complete tlbies and TLB loads in the third cycle
1363 sync
+= r1
.mmu_done
.eq(r0_valid
& (r0
.tlbie | r0
.tlbld
))
1365 with m
.If((req_op
== Op
.OP_LOAD_HIT
) |
(req_op
== Op
.OP_STCX_FAIL
)):
1366 with m
.If(~r0
.mmu_req
):
1367 sync
+= r1
.ls_valid
.eq(1)
1369 sync
+= r1
.mmu_done
.eq(1)
1371 with m
.If(r1
.write_tag
):
1372 # Store new tag in selected way
1373 replace_way_onehot
= Signal(NUM_WAYS
)
1374 comb
+= replace_way_onehot
.eq(1<<replace_way
)
1375 for i
in range(NUM_WAYS
):
1376 with m
.If(replace_way_onehot
[i
]):
1377 ct
= Signal(TAG_RAM_WIDTH
)
1378 comb
+= ct
.eq(cache_tags
[r1
.store_index
].tag
)
1379 comb
+= ct
.word_select(i
, TAG_WIDTH
).eq(r1
.reload_tag
)
1380 sync
+= cache_tags
[r1
.store_index
].tag
.eq(ct
)
1381 sync
+= r1
.store_way
.eq(replace_way
)
1382 sync
+= r1
.write_tag
.eq(0)
1384 # Take request from r1.req if there is one there,
1385 # else from req_op, ra, etc.
1387 comb
+= req
.eq(r1
.req
)
1389 comb
+= req
.op
.eq(req_op
)
1390 comb
+= req
.valid
.eq(req_go
)
1391 comb
+= req
.mmu_req
.eq(r0
.mmu_req
)
1392 comb
+= req
.dcbz
.eq(r0
.req
.dcbz
)
1393 comb
+= req
.real_addr
.eq(ra
)
1395 with m
.If(r0
.req
.dcbz
):
1396 # force data to 0 for dcbz
1397 comb
+= req
.data
.eq(0)
1398 with m
.Elif(r0
.d_valid
):
1399 comb
+= req
.data
.eq(r0
.req
.data
)
1401 comb
+= req
.data
.eq(d_in
.data
)
1403 # Select all bytes for dcbz
1404 # and for cacheable loads
1405 with m
.If(r0
.req
.dcbz |
(r0
.req
.load
& ~r0
.req
.nc
)):
1406 comb
+= req
.byte_sel
.eq(~
0) # all 1s
1408 comb
+= req
.byte_sel
.eq(r0
.req
.byte_sel
)
1409 comb
+= req
.hit_way
.eq(req_hit_way
)
1410 comb
+= req
.same_tag
.eq(req_same_tag
)
1412 # Store the incoming request from r0,
1413 # if it is a slow request
1414 # Note that r1.full = 1 implies req_op = OP_NONE
1415 with m
.If((req_op
== Op
.OP_LOAD_MISS
)
1416 |
(req_op
== Op
.OP_LOAD_NC
)
1417 |
(req_op
== Op
.OP_STORE_MISS
)
1418 |
(req_op
== Op
.OP_STORE_HIT
)):
1419 sync
+= r1
.req
.eq(req
)
1420 sync
+= r1
.full
.eq(1)
1422 # Main state machine
1423 with m
.Switch(r1
.state
):
1425 with m
.Case(State
.IDLE
):
1426 sync
+= r1
.wb
.adr
.eq(req
.real_addr
[ROW_LINE_BITS
:])
1427 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1428 sync
+= r1
.wb
.dat
.eq(req
.data
)
1429 sync
+= r1
.dcbz
.eq(req
.dcbz
)
1431 # Keep track of our index and way
1432 # for subsequent stores.
1433 sync
+= r1
.store_index
.eq(req_idx
)
1434 sync
+= r1
.store_row
.eq(req_row
)
1435 sync
+= r1
.end_row_ix
.eq(get_row_of_line(req_row
)-1)
1436 sync
+= r1
.reload_tag
.eq(req_tag
)
1437 sync
+= r1
.req
.same_tag
.eq(1)
1439 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1440 sync
+= r1
.store_way
.eq(req
.hit_way
)
1442 # Reset per-row valid bits,
1443 # ready for handling OP_LOAD_MISS
1444 for i
in range(ROW_PER_LINE
):
1445 sync
+= r1
.rows_valid
[i
].eq(0)
1447 with m
.If(req_op
!= Op
.OP_NONE
):
1448 sync
+= Display("cache op %d", req
.op
)
1450 with m
.Switch(req
.op
):
1451 with m
.Case(Op
.OP_LOAD_HIT
):
1452 # stay in IDLE state
1455 with m
.Case(Op
.OP_LOAD_MISS
):
1456 sync
+= Display("cache miss real addr: %x " \
1458 req
.real_addr
, req_row
, req_tag
)
1460 # Start the wishbone cycle
1461 sync
+= r1
.wb
.we
.eq(0)
1462 sync
+= r1
.wb
.cyc
.eq(1)
1463 sync
+= r1
.wb
.stb
.eq(1)
1465 # Track that we had one request sent
1466 sync
+= r1
.state
.eq(State
.RELOAD_WAIT_ACK
)
1467 sync
+= r1
.write_tag
.eq(1)
1469 with m
.Case(Op
.OP_LOAD_NC
):
1470 sync
+= r1
.wb
.cyc
.eq(1)
1471 sync
+= r1
.wb
.stb
.eq(1)
1472 sync
+= r1
.wb
.we
.eq(0)
1473 sync
+= r1
.state
.eq(State
.NC_LOAD_WAIT_ACK
)
1475 with m
.Case(Op
.OP_STORE_HIT
, Op
.OP_STORE_MISS
):
1476 with m
.If(~req
.dcbz
):
1477 sync
+= r1
.state
.eq(State
.STORE_WAIT_ACK
)
1478 sync
+= r1
.acks_pending
.eq(1)
1479 sync
+= r1
.full
.eq(0)
1480 sync
+= r1
.slow_valid
.eq(1)
1482 with m
.If(~req
.mmu_req
):
1483 sync
+= r1
.ls_valid
.eq(1)
1485 sync
+= r1
.mmu_done
.eq(1)
1487 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1488 sync
+= r1
.write_bram
.eq(1)
1490 # dcbz is handled much like a load miss except
1491 # that we are writing to memory instead of reading
1492 sync
+= r1
.state
.eq(State
.RELOAD_WAIT_ACK
)
1494 with m
.If(req
.op
== Op
.OP_STORE_MISS
):
1495 sync
+= r1
.write_tag
.eq(1)
1497 sync
+= r1
.wb
.we
.eq(1)
1498 sync
+= r1
.wb
.cyc
.eq(1)
1499 sync
+= r1
.wb
.stb
.eq(1)
1501 # OP_NONE and OP_BAD do nothing
1502 # OP_BAD & OP_STCX_FAIL were
1503 # handled above already
1504 with m
.Case(Op
.OP_NONE
):
1506 with m
.Case(Op
.OP_BAD
):
1508 with m
.Case(Op
.OP_STCX_FAIL
):
1511 with m
.Case(State
.RELOAD_WAIT_ACK
):
1512 ld_stbs_done
= Signal()
1513 # Requests are all sent if stb is 0
1514 comb
+= ld_stbs_done
.eq(~r1
.wb
.stb
)
1516 # If we are still sending requests, was one accepted?
1517 with m
.If((~bus
.stall
) & r1
.wb
.stb
):
1518 # That was the last word? We are done sending.
1519 # Clear stb and set ld_stbs_done so we can handle an
1520 # eventual last ack on the same cycle.
1521 # sigh - reconstruct wb adr with 3 extra 0s at front
1522 wb_adr
= Cat(Const(0, ROW_OFF_BITS
), r1
.wb
.adr
)
1523 with m
.If(is_last_row_addr(wb_adr
, r1
.end_row_ix
)):
1524 sync
+= r1
.wb
.stb
.eq(0)
1525 comb
+= ld_stbs_done
.eq(1)
1527 # Calculate the next row address in the current cache line
1528 row
= Signal(LINE_OFF_BITS
-ROW_OFF_BITS
)
1529 comb
+= row
.eq(r1
.wb
.adr
)
1530 sync
+= r1
.wb
.adr
[:LINE_OFF_BITS
-ROW_OFF_BITS
].eq(row
+1)
1532 # Incoming acks processing
1533 sync
+= r1
.forward_valid1
.eq(bus
.ack
)
1535 srow
= Signal(ROW_LINE_BITS
)
1536 comb
+= srow
.eq(r1
.store_row
)
1537 sync
+= r1
.rows_valid
[srow
].eq(1)
1539 # If this is the data we were looking for,
1540 # we can complete the request next cycle.
1541 # Compare the whole address in case the
1542 # request in r1.req is not the one that
1543 # started this refill.
1544 with m
.If(req
.valid
& r1
.req
.same_tag
&
1545 ((r1
.dcbz
& r1
.req
.dcbz
) |
1546 (~r1
.dcbz
& (r1
.req
.op
== Op
.OP_LOAD_MISS
))) &
1547 (r1
.store_row
== get_row(req
.real_addr
))):
1548 sync
+= r1
.full
.eq(0)
1549 sync
+= r1
.slow_valid
.eq(1)
1550 with m
.If(~r1
.mmu_req
):
1551 sync
+= r1
.ls_valid
.eq(1)
1553 sync
+= r1
.mmu_done
.eq(1)
1554 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1555 sync
+= r1
.use_forward1
.eq(1)
1557 # Check for completion
1558 with m
.If(ld_stbs_done
& is_last_row(r1
.store_row
,
1560 # Complete wishbone cycle
1561 sync
+= r1
.wb
.cyc
.eq(0)
1563 # Cache line is now valid
1564 cv
= Signal(INDEX_BITS
)
1565 comb
+= cv
.eq(cache_tags
[r1
.store_index
].valid
)
1566 comb
+= cv
.bit_select(r1
.store_way
, 1).eq(1)
1567 sync
+= cache_tags
[r1
.store_index
].valid
.eq(cv
)
1569 sync
+= r1
.state
.eq(State
.IDLE
)
1570 sync
+= Display("cache valid set %x "
1572 cv
, r1
.store_index
, r1
.store_way
)
1574 # Increment store row counter
1575 sync
+= r1
.store_row
.eq(next_row(r1
.store_row
))
1577 with m
.Case(State
.STORE_WAIT_ACK
):
1578 st_stbs_done
= Signal()
1580 adjust_acks
= Signal(3)
1582 comb
+= st_stbs_done
.eq(~r1
.wb
.stb
)
1583 comb
+= acks
.eq(r1
.acks_pending
)
1585 with m
.If(r1
.inc_acks
!= r1
.dec_acks
):
1586 with m
.If(r1
.inc_acks
):
1587 comb
+= adjust_acks
.eq(acks
+ 1)
1589 comb
+= adjust_acks
.eq(acks
- 1)
1591 comb
+= adjust_acks
.eq(acks
)
1593 sync
+= r1
.acks_pending
.eq(adjust_acks
)
1595 # Clear stb when slave accepted request
1596 with m
.If(~bus
.stall
):
1597 # See if there is another store waiting
1598 # to be done which is in the same real page.
1599 with m
.If(req
.valid
):
1600 _ra
= req
.real_addr
[ROW_LINE_BITS
:SET_SIZE_BITS
]
1601 sync
+= r1
.wb
.adr
[0:SET_SIZE_BITS
].eq(_ra
)
1602 sync
+= r1
.wb
.dat
.eq(req
.data
)
1603 sync
+= r1
.wb
.sel
.eq(req
.byte_sel
)
1605 with m
.If((adjust_acks
< 7) & req
.same_tag
&
1606 ((req
.op
== Op
.OP_STORE_MISS
)
1607 |
(req
.op
== Op
.OP_STORE_HIT
))):
1608 sync
+= r1
.wb
.stb
.eq(1)
1609 comb
+= st_stbs_done
.eq(0)
1611 with m
.If(req
.op
== Op
.OP_STORE_HIT
):
1612 sync
+= r1
.write_bram
.eq(1)
1613 sync
+= r1
.full
.eq(0)
1614 sync
+= r1
.slow_valid
.eq(1)
1616 # Store requests never come from the MMU
1617 sync
+= r1
.ls_valid
.eq(1)
1618 comb
+= st_stbs_done
.eq(0)
1619 sync
+= r1
.inc_acks
.eq(1)
1621 sync
+= r1
.wb
.stb
.eq(0)
1622 comb
+= st_stbs_done
.eq(1)
1624 # Got ack ? See if complete.
1626 with m
.If(st_stbs_done
& (adjust_acks
== 1)):
1627 sync
+= r1
.state
.eq(State
.IDLE
)
1628 sync
+= r1
.wb
.cyc
.eq(0)
1629 sync
+= r1
.wb
.stb
.eq(0)
1630 sync
+= r1
.dec_acks
.eq(1)
1632 with m
.Case(State
.NC_LOAD_WAIT_ACK
):
1633 # Clear stb when slave accepted request
1634 with m
.If(~bus
.stall
):
1635 sync
+= r1
.wb
.stb
.eq(0)
1637 # Got ack ? complete.
1639 sync
+= r1
.state
.eq(State
.IDLE
)
1640 sync
+= r1
.full
.eq(0)
1641 sync
+= r1
.slow_valid
.eq(1)
1643 with m
.If(~r1
.mmu_req
):
1644 sync
+= r1
.ls_valid
.eq(1)
1646 sync
+= r1
.mmu_done
.eq(1)
1648 sync
+= r1
.forward_sel
.eq(~
0) # all 1s
1649 sync
+= r1
.use_forward1
.eq(1)
1650 sync
+= r1
.wb
.cyc
.eq(0)
1651 sync
+= r1
.wb
.stb
.eq(0)
1653 def dcache_log(self
, m
, r1
, valid_ra
, tlb_hit
, stall_out
):
1656 d_out
, bus
, log_out
= self
.d_out
, self
.bus
, self
.log_out
1658 sync
+= log_out
.eq(Cat(r1
.state
[:3], valid_ra
, tlb_hit
.way
[:3],
1659 stall_out
, req_op
[:3], d_out
.valid
, d_out
.error
,
1660 r1
.wb
.cyc
, r1
.wb
.stb
, bus
.ack
, bus
.stall
,
1663 def elaborate(self
, platform
):
1669 # Storage. Hopefully "cache_rows" is a BRAM, the rest is LUTs
1670 cache_tags
= CacheTagArray()
1671 cache_tag_set
= Signal(TAG_RAM_WIDTH
)
1673 # TODO attribute ram_style : string;
1674 # TODO attribute ram_style of cache_tags : signal is "distributed";
1676 """note: these are passed to nmigen.hdl.Memory as "attributes".
1677 don't know how, just that they are.
1679 # TODO attribute ram_style of
1680 # dtlb_tags : signal is "distributed";
1681 # TODO attribute ram_style of
1682 # dtlb_ptes : signal is "distributed";
1684 r0
= RegStage0("r0")
1687 r1
= RegStage1("r1")
1689 reservation
= Reservation()
1691 # Async signals on incoming request
1692 req_index
= Signal(INDEX_BITS
)
1693 req_row
= Signal(ROW_BITS
)
1694 req_hit_way
= Signal(WAY_BITS
)
1695 req_tag
= Signal(TAG_BITS
)
1697 req_data
= Signal(64)
1698 req_same_tag
= Signal()
1701 early_req_row
= Signal(ROW_BITS
)
1703 cancel_store
= Signal()
1705 clear_rsrv
= Signal()
1710 use_forward1_next
= Signal()
1711 use_forward2_next
= Signal()
1713 cache_out_row
= Signal(WB_DATA_BITS
)
1715 plru_victim
= Signal(WAY_BITS
)
1716 replace_way
= Signal(WAY_BITS
)
1718 # Wishbone read/write/cache write formatting signals
1722 tlb_way
= TLBRecord("tlb_way")
1723 tlb_req_index
= Signal(TLB_SET_BITS
)
1724 tlb_hit
= TLBHit("tlb_hit")
1725 pte
= Signal(TLB_PTE_BITS
)
1726 ra
= Signal(REAL_ADDR_BITS
)
1728 perm_attr
= PermAttr("dc_perms")
1731 access_ok
= Signal()
1733 tlb_plru_victim
= Signal(TLB_WAY_BITS
)
1735 # we don't yet handle collisions between loadstore1 requests
1737 comb
+= self
.m_out
.stall
.eq(0)
1739 # Hold off the request in r0 when r1 has an uncompleted request
1740 comb
+= r0_stall
.eq(r0_full
& (r1
.full | d_in
.hold
))
1741 comb
+= r0_valid
.eq(r0_full
& ~r1
.full
& ~d_in
.hold
)
1742 comb
+= self
.stall_out
.eq(r0_stall
)
1744 # deal with litex not doing wishbone pipeline mode
1745 # XXX in wrong way. FIFOs are needed in the SRAM test
1746 # so that stb/ack match up. same thing done in icache.py
1747 comb
+= self
.bus
.stall
.eq(self
.bus
.cyc
& ~self
.bus
.ack
)
1749 # Wire up wishbone request latch out of stage 1
1750 comb
+= self
.bus
.we
.eq(r1
.wb
.we
)
1751 comb
+= self
.bus
.adr
.eq(r1
.wb
.adr
)
1752 comb
+= self
.bus
.sel
.eq(r1
.wb
.sel
)
1753 comb
+= self
.bus
.stb
.eq(r1
.wb
.stb
)
1754 comb
+= self
.bus
.dat_w
.eq(r1
.wb
.dat
)
1755 comb
+= self
.bus
.cyc
.eq(r1
.wb
.cyc
)
1757 # create submodule TLBUpdate
1758 m
.submodules
.dtlb_update
= self
.dtlb_update
= DTLBUpdate()
1760 # call sub-functions putting everything together, using shared
1761 # signals established above
1762 self
.stage_0(m
, r0
, r1
, r0_full
)
1763 self
.tlb_read(m
, r0_stall
, tlb_way
)
1764 self
.tlb_search(m
, tlb_req_index
, r0
, r0_valid
,
1766 pte
, tlb_hit
, valid_ra
, perm_attr
, ra
)
1767 self
.tlb_update(m
, r0_valid
, r0
, tlb_req_index
,
1768 tlb_hit
, tlb_plru_victim
,
1770 self
.maybe_plrus(m
, r1
, plru_victim
)
1771 self
.maybe_tlb_plrus(m
, r1
, tlb_plru_victim
, tlb_req_index
)
1772 self
.cache_tag_read(m
, r0_stall
, req_index
, cache_tag_set
, cache_tags
)
1773 self
.dcache_request(m
, r0
, ra
, req_index
, req_row
, req_tag
,
1774 r0_valid
, r1
, cache_tags
, replace_way
,
1775 use_forward1_next
, use_forward2_next
,
1776 req_hit_way
, plru_victim
, rc_ok
, perm_attr
,
1777 valid_ra
, perm_ok
, access_ok
, req_op
, req_go
,
1778 tlb_hit
, tlb_way
, cache_tag_set
,
1779 cancel_store
, req_same_tag
, r0_stall
, early_req_row
)
1780 self
.reservation_comb(m
, cancel_store
, set_rsrv
, clear_rsrv
,
1781 r0_valid
, r0
, reservation
)
1782 self
.reservation_reg(m
, r0_valid
, access_ok
, set_rsrv
, clear_rsrv
,
1784 self
.writeback_control(m
, r1
, cache_out_row
)
1785 self
.rams(m
, r1
, early_req_row
, cache_out_row
, replace_way
)
1786 self
.dcache_fast_hit(m
, req_op
, r0_valid
, r0
, r1
,
1787 req_hit_way
, req_index
, req_tag
, access_ok
,
1788 tlb_hit
, tlb_req_index
)
1789 self
.dcache_slow(m
, r1
, use_forward1_next
, use_forward2_next
,
1791 req_hit_way
, req_same_tag
,
1792 r0_valid
, req_op
, cache_tags
, req_go
, ra
)
1793 #self.dcache_log(m, r1, valid_ra, tlb_hit, stall_out)
1798 if __name__
== '__main__':
1800 vl
= rtlil
.convert(dut
, ports
=[])
1801 with
open("test_dcache.il", "w") as f
: