3 based on microwatt loadstore1.vhdl, but conforming to PortInterface.
4 unlike loadstore1.vhdl this does *not* deal with actual Load/Store
5 ops: that job is handled by LDSTCompUnit, which talks to LoadStore1
6 by way of PortInterface. PortInterface is where things need extending,
7 such as adding dcbz support, etc.
9 this module basically handles "pure" load / store operations, and
10 its first job is to ask the D-Cache for the data. if that fails,
11 the second task (if virtual memory is enabled) is to ask the MMU
12 to perform a TLB, then to go *back* to the cache and ask again.
16 * https://bugs.libre-soc.org/show_bug.cgi?id=465
20 from nmigen
import (Elaboratable
, Module
, Signal
, Shape
, unsigned
, Cat
, Mux
,
23 from nmutil
.iocontrol
import RecordObject
24 from nmutil
.util
import rising_edge
, Display
25 from enum
import Enum
, unique
27 from soc
.experiment
.dcache
import DCache
28 from soc
.experiment
.pimem
import PortInterfaceBase
29 from soc
.experiment
.mem_types
import LoadStore1ToMMUType
30 from soc
.experiment
.mem_types
import MMUToLoadStore1Type
32 from soc
.minerva
.wishbone
import make_wb_layout
33 from soc
.bus
.sram
import SRAM
34 from nmutil
.util
import Display
39 IDLE
= 0 # ready for instruction
40 ACK_WAIT
= 1 # waiting for ack from dcache
41 MMU_LOOKUP
= 2 # waiting for MMU to look up translation
42 TLBIE_WAIT
= 3 # waiting for MMU to finish doing a tlbie
45 # captures the LDSTRequest from the PortInterface, which "blips" most
46 # of this at us (pipeline-style).
47 class LDSTRequest(RecordObject
):
48 def __init__(self
, name
=None):
49 RecordObject
.__init
__(self
, name
=name
)
53 self
.addr
= Signal(64)
54 # self.store_data = Signal(64) # this is already sync (on a delay)
55 self
.byte_sel
= Signal(8)
56 self
.nc
= Signal() # non-cacheable access
57 self
.virt_mode
= Signal()
58 self
.priv_mode
= Signal()
59 self
.align_intr
= Signal()
61 # glue logic for microwatt mmu and dcache
62 class LoadStore1(PortInterfaceBase
):
63 def __init__(self
, pspec
):
65 self
.disable_cache
= (hasattr(pspec
, "disable_cache") and
66 pspec
.disable_cache
== True)
67 regwid
= pspec
.reg_wid
68 addrwid
= pspec
.addr_wid
70 super().__init
__(regwid
, addrwid
)
71 self
.dcache
= DCache()
72 # these names are from the perspective of here (LoadStore1)
73 self
.d_out
= self
.dcache
.d_in
# in to dcache is out for LoadStore
74 self
.d_in
= self
.dcache
.d_out
# out from dcache is in for LoadStore
75 self
.m_out
= LoadStore1ToMMUType() # out *to* MMU
76 self
.m_in
= MMUToLoadStore1Type() # in *from* MMU
77 self
.req
= LDSTRequest(name
="ldst_req")
79 # TODO, convert dcache wb_in/wb_out to "standard" nmigen Wishbone bus
80 self
.dbus
= Record(make_wb_layout(pspec
))
82 # for creating a single clock blip to DCache
83 self
.d_valid
= Signal()
84 self
.d_w_valid
= Signal()
85 self
.d_validblip
= Signal()
87 # DSISR and DAR cached values. note that the MMU FSM is where
88 # these are accessed by OP_MTSPR/OP_MFSPR, on behalf of LoadStore1.
89 # by contrast microwatt has the spr set/get done *in* loadstore1.vhdl
90 self
.dsisr
= Signal(64)
93 # state info for LD/ST
95 # latch most of the input request
99 self
.addr
= Signal(64)
100 self
.store_data
= Signal(64)
101 self
.load_data
= Signal(64)
102 self
.byte_sel
= Signal(8)
103 #self.xerc : xer_common_t;
104 #self.reserve = Signal()
105 #self.atomic = Signal()
106 #self.atomic_last = Signal()
108 self
.nc
= Signal() # non-cacheable access
109 self
.virt_mode
= Signal()
110 self
.priv_mode
= Signal()
111 self
.state
= Signal(State
)
112 self
.instr_fault
= Signal()
113 self
.align_intr
= Signal()
115 self
.wait_dcache
= Signal()
116 self
.wait_mmu
= Signal()
117 #self.mode_32bit = Signal()
118 #self.intr_vec : integer range 0 to 16#fff#;
119 #self.nia = Signal(64)
120 #self.srr1 = Signal(16)
122 def set_wr_addr(self
, m
, addr
, mask
, misalign
, msr_pr
):
123 m
.d
.comb
+= self
.req
.load
.eq(0) # store operation
124 m
.d
.comb
+= self
.req
.byte_sel
.eq(mask
)
125 m
.d
.comb
+= self
.req
.addr
.eq(addr
)
126 m
.d
.comb
+= self
.req
.priv_mode
.eq(~msr_pr
) # not-problem ==> priv
127 m
.d
.comb
+= self
.req
.virt_mode
.eq(msr_pr
) # problem-state ==> virt
128 m
.d
.comb
+= self
.req
.align_intr
.eq(misalign
)
130 dcbz
= self
.pi
.is_dcbz
132 m
.d
.comb
+= Display("set_wr_addr: is_dcbz")
133 m
.d
.comb
+= self
.req
.dcbz
.eq(dcbz
)
135 # option to disable the cache entirely for write
136 if self
.disable_cache
:
137 m
.d
.comb
+= self
.req
.nc
.eq(1)
140 def set_rd_addr(self
, m
, addr
, mask
, misalign
, msr_pr
):
141 m
.d
.comb
+= self
.d_valid
.eq(1)
142 m
.d
.comb
+= self
.req
.load
.eq(1) # load operation
143 m
.d
.comb
+= self
.req
.byte_sel
.eq(mask
)
144 m
.d
.comb
+= self
.req
.align_intr
.eq(misalign
)
145 m
.d
.comb
+= self
.req
.addr
.eq(addr
)
146 m
.d
.comb
+= self
.req
.priv_mode
.eq(~msr_pr
) # not-problem ==> priv
147 m
.d
.comb
+= self
.req
.virt_mode
.eq(msr_pr
) # problem-state ==> virt
148 # BAD HACK! disable cacheing on LD when address is 0xCxxx_xxxx
149 # this is for peripherals. same thing done in Microwatt loadstore1.vhdl
150 with m
.If(addr
[28:] == Const(0xc, 4)):
151 m
.d
.comb
+= self
.req
.nc
.eq(1)
152 # option to disable the cache entirely for read
153 if self
.disable_cache
:
154 m
.d
.comb
+= self
.req
.nc
.eq(1)
155 return None #FIXME return value
157 def set_wr_data(self
, m
, data
, wen
):
158 # do the "blip" on write data
159 m
.d
.comb
+= self
.d_valid
.eq(1)
160 # put data into comb which is picked up in main elaborate()
161 m
.d
.comb
+= self
.d_w_valid
.eq(1)
162 m
.d
.comb
+= self
.store_data
.eq(data
)
163 #m.d.sync += self.d_out.byte_sel.eq(wen) # this might not be needed
164 st_ok
= self
.done
# TODO indicates write data is valid
167 def get_rd_data(self
, m
):
168 ld_ok
= self
.done
# indicates read data is valid
169 data
= self
.load_data
# actual read data
172 def elaborate(self
, platform
):
173 m
= super().elaborate(platform
)
174 comb
, sync
= m
.d
.comb
, m
.d
.sync
176 # create dcache module
177 m
.submodules
.dcache
= dcache
= self
.dcache
180 d_out
, d_in
, dbus
= self
.d_out
, self
.d_in
, self
.dbus
181 m_out
, m_in
= self
.m_out
, self
.m_in
183 exception
= exc
.happened
186 # copy of address, but gets over-ridden for OP_FETCH_FAILED
188 m
.d
.comb
+= maddr
.eq(self
.addr
)
190 # create a blip (single pulse) on valid read/write request
191 # this can be over-ridden in the FSM to get dcache to re-run
192 # a request when MMU_LOOKUP completes.
193 m
.d
.comb
+= self
.d_validblip
.eq(rising_edge(m
, self
.d_valid
))
194 ldst_r
= LDSTRequest("ldst_r")
197 with m
.Switch(self
.state
):
198 with m
.Case(State
.IDLE
):
199 with m
.If(self
.d_validblip
& ~exc
.happened
):
200 comb
+= self
.busy
.eq(1)
201 sync
+= self
.state
.eq(State
.ACK_WAIT
)
202 sync
+= ldst_r
.eq(self
.req
) # copy of LDSTRequest on "blip"
206 # waiting for completion
207 with m
.Case(State
.ACK_WAIT
):
208 comb
+= self
.busy
.eq(~exc
.happened
)
210 with m
.If(d_in
.error
):
211 # cache error is not necessarily "final", it could
212 # be that it was just a TLB miss
213 with m
.If(d_in
.cache_paradox
):
214 comb
+= exception
.eq(1)
215 sync
+= self
.state
.eq(State
.IDLE
)
217 sync
+= self
.dsisr
[63 - 38].eq(~self
.load
)
218 # XXX there is no architected bit for this
219 # (probably should be a machine check in fact)
220 sync
+= self
.dsisr
[63 - 35].eq(d_in
.cache_paradox
)
223 # Look up the translation for TLB miss
224 # and also for permission error and RC error
225 # in case the PTE has been updated.
227 sync
+= self
.state
.eq(State
.MMU_LOOKUP
)
228 with m
.If(d_in
.valid
):
229 m
.d
.comb
+= self
.done
.eq(~mmureq
) # done if not doing MMU
230 with m
.If(self
.done
):
231 sync
+= Display("ACK_WAIT, done %x", self
.addr
)
232 sync
+= self
.state
.eq(State
.IDLE
)
234 with m
.If(self
.load
):
235 m
.d
.comb
+= self
.load_data
.eq(d_in
.data
)
237 # waiting here for the MMU TLB lookup to complete.
238 # either re-try the dcache lookup or throw MMU exception
239 with m
.Case(State
.MMU_LOOKUP
):
240 comb
+= self
.busy
.eq(1)
241 with m
.If(m_in
.done
):
242 with m
.If(~self
.instr_fault
):
243 sync
+= Display("MMU_LOOKUP, done %x -> %x",
244 self
.addr
, d_out
.addr
)
245 # retry the request now that the MMU has
246 # installed a TLB entry, if not exception raised
247 m
.d
.comb
+= self
.d_out
.valid
.eq(~exception
)
248 sync
+= self
.state
.eq(State
.ACK_WAIT
)
251 sync
+= Display("MMU_LOOKUP, exception %x", self
.addr
)
252 # instruction lookup fault: store address in DAR
253 comb
+= exc
.happened
.eq(1)
254 sync
+= self
.dar
.eq(self
.addr
)
257 # MMU RADIX exception thrown
258 comb
+= exception
.eq(1)
259 sync
+= self
.dsisr
[63 - 33].eq(m_in
.invalid
)
260 sync
+= self
.dsisr
[63 - 36].eq(m_in
.perm_error
)
261 sync
+= self
.dsisr
[63 - 38].eq(self
.load
)
262 sync
+= self
.dsisr
[63 - 44].eq(m_in
.badtree
)
263 sync
+= self
.dsisr
[63 - 45].eq(m_in
.rc_error
)
265 with m
.Case(State
.TLBIE_WAIT
):
268 # alignment error: store address in DAR
269 with m
.If(self
.align_intr
):
270 comb
+= exc
.happened
.eq(1)
271 sync
+= self
.dar
.eq(self
.addr
)
273 # happened, alignment, instr_fault, invalid.
274 # note that all of these flow through - eventually to the TRAP
275 # pipeline, via PowerDecoder2.
276 comb
+= exc
.invalid
.eq(m_in
.invalid
)
277 comb
+= exc
.alignment
.eq(self
.align_intr
)
278 comb
+= exc
.instr_fault
.eq(self
.instr_fault
)
279 # badtree, perm_error, rc_error, segment_fault
280 comb
+= exc
.badtree
.eq(m_in
.badtree
)
281 comb
+= exc
.perm_error
.eq(m_in
.perm_error
)
282 comb
+= exc
.rc_error
.eq(m_in
.rc_error
)
283 comb
+= exc
.segment_fault
.eq(m_in
.segerr
)
285 # TODO, connect dcache wb_in/wb_out to "standard" nmigen Wishbone bus
286 comb
+= dbus
.adr
.eq(dcache
.wb_out
.adr
)
287 comb
+= dbus
.dat_w
.eq(dcache
.wb_out
.dat
)
288 comb
+= dbus
.sel
.eq(dcache
.wb_out
.sel
)
289 comb
+= dbus
.cyc
.eq(dcache
.wb_out
.cyc
)
290 comb
+= dbus
.stb
.eq(dcache
.wb_out
.stb
)
291 comb
+= dbus
.we
.eq(dcache
.wb_out
.we
)
293 comb
+= dcache
.wb_in
.dat
.eq(dbus
.dat_r
)
294 comb
+= dcache
.wb_in
.ack
.eq(dbus
.ack
)
295 if hasattr(dbus
, "stall"):
296 comb
+= dcache
.wb_in
.stall
.eq(dbus
.stall
)
298 # update out d data when flag set
299 with m
.If(self
.d_w_valid
):
300 m
.d
.sync
+= d_out
.data
.eq(self
.store_data
)
302 # m.d.sync += d_out.data.eq(0)
303 # unit test passes with that change
305 # this must move into the FSM, conditionally noticing that
306 # the "blip" comes from self.d_validblip.
307 # task 1: look up in dcache
308 # task 2: if dcache fails, look up in MMU.
309 # do **NOT** confuse the two.
310 with m
.If(self
.d_validblip
):
311 m
.d
.comb
+= self
.d_out
.valid
.eq(~exc
.happened
)
312 m
.d
.comb
+= d_out
.load
.eq(self
.req
.load
)
313 m
.d
.comb
+= d_out
.byte_sel
.eq(self
.req
.byte_sel
)
314 m
.d
.comb
+= self
.addr
.eq(self
.req
.addr
)
315 m
.d
.comb
+= d_out
.nc
.eq(self
.req
.nc
)
316 m
.d
.comb
+= d_out
.priv_mode
.eq(self
.req
.priv_mode
)
317 m
.d
.comb
+= d_out
.virt_mode
.eq(self
.req
.virt_mode
)
318 m
.d
.comb
+= self
.align_intr
.eq(self
.req
.align_intr
)
319 #m.d.comb += Display("validblip dcbz=%i addr=%x",self.req.dcbz,self.req.addr)
320 m
.d
.comb
+= d_out
.dcbz
.eq(self
.req
.dcbz
)
322 m
.d
.comb
+= d_out
.load
.eq(ldst_r
.load
)
323 m
.d
.comb
+= d_out
.byte_sel
.eq(ldst_r
.byte_sel
)
324 m
.d
.comb
+= self
.addr
.eq(ldst_r
.addr
)
325 m
.d
.comb
+= d_out
.nc
.eq(ldst_r
.nc
)
326 m
.d
.comb
+= d_out
.priv_mode
.eq(ldst_r
.priv_mode
)
327 m
.d
.comb
+= d_out
.virt_mode
.eq(ldst_r
.virt_mode
)
328 m
.d
.comb
+= self
.align_intr
.eq(ldst_r
.align_intr
)
329 #m.d.comb += Display("no_validblip dcbz=%i addr=%x",ldst_r.dcbz,ldst_r.addr)
330 m
.d
.comb
+= d_out
.dcbz
.eq(ldst_r
.dcbz
)
332 # XXX these should be possible to remove but for some reason
333 # cannot be... yet. TODO, investigate
334 m
.d
.comb
+= self
.load_data
.eq(d_in
.data
)
335 m
.d
.comb
+= d_out
.addr
.eq(self
.addr
)
337 # Update outputs to MMU
338 m
.d
.comb
+= m_out
.valid
.eq(mmureq
)
339 m
.d
.comb
+= m_out
.iside
.eq(self
.instr_fault
)
340 m
.d
.comb
+= m_out
.load
.eq(ldst_r
.load
)
341 # m_out.priv <= r.priv_mode; TODO
342 m
.d
.comb
+= m_out
.tlbie
.eq(self
.tlbie
)
343 # m_out.mtspr <= mmu_mtspr; # TODO
344 # m_out.sprn <= sprn; # TODO
345 m
.d
.comb
+= m_out
.addr
.eq(maddr
)
346 # m_out.slbia <= l_in.insn(7); # TODO: no idea what this is
347 # m_out.rs <= l_in.data; # nope, probably not needed, TODO investigate
352 yield from super().ports()
356 class TestSRAMLoadStore1(LoadStore1
):
357 def __init__(self
, pspec
):
358 super().__init
__(pspec
)
360 # small 32-entry Memory
361 if (hasattr(pspec
, "dmem_test_depth") and
362 isinstance(pspec
.dmem_test_depth
, int)):
363 depth
= pspec
.dmem_test_depth
366 print("TestSRAMBareLoadStoreUnit depth", depth
)
368 self
.mem
= Memory(width
=pspec
.reg_wid
, depth
=depth
)
370 def elaborate(self
, platform
):
371 m
= super().elaborate(platform
)
373 m
.submodules
.sram
= sram
= SRAM(memory
=self
.mem
, granularity
=8,
374 features
={'cti', 'bte', 'err'})
377 # directly connect the wishbone bus of LoadStoreUnitInterface to SRAM
378 # note: SRAM is a target (slave), dbus is initiator (master)
379 fanouts
= ['dat_w', 'sel', 'cyc', 'stb', 'we', 'cti', 'bte']
380 fanins
= ['dat_r', 'ack', 'err']
381 for fanout
in fanouts
:
382 print("fanout", fanout
, getattr(sram
.bus
, fanout
).shape(),
383 getattr(dbus
, fanout
).shape())
384 comb
+= getattr(sram
.bus
, fanout
).eq(getattr(dbus
, fanout
))
385 comb
+= getattr(sram
.bus
, fanout
).eq(getattr(dbus
, fanout
))
387 comb
+= getattr(dbus
, fanin
).eq(getattr(sram
.bus
, fanin
))
389 comb
+= sram
.bus
.adr
.eq(dbus
.adr
)