1 """LOAD / STORE Computation Unit.
3 This module covers POWER9-compliant Load and Store operations,
4 with selection on each between immediate and indexed mode as
5 options for the calculation of the Effective Address (EA),
6 and also "update" mode which optionally stores that EA into
7 an additional register.
10 Note: it took 15 attempts over several weeks to redraw the diagram
11 needed to capture this FSM properly. To understand it fully, please
12 take the time to review the links, video, and diagram.
15 Stores are activated when Go_Store is enabled, and use a sync'd "ADD" to
16 compute the "Effective Address", and, when ready the operand (src3_i)
17 is stored in the computed address (passed through to the PortInterface)
19 Loads are activated when Go_Write[0] is enabled. The EA is computed,
20 and (as long as there was no exception) the data comes out (at any
21 time from the PortInterface), and is captured by the LDCompSTUnit.
23 Both LD and ST may request that the address be computed from summing
24 operand1 (src[0]) with operand2 (src[1]) *or* by summing operand1 with
25 the immediate (from the opcode).
27 Both LD and ST may also request "update" mode (op_is_update) which
28 activates the use of Go_Write[1] to control storage of the EA into
29 a *second* operand in the register file.
31 Thus this module has *TWO* write-requests to the register file and
32 *THREE* read-requests to the register file (not all at the same time!)
33 The regfile port usage is:
45 It's a multi-level Finite State Machine that (unfortunately) nmigen.FSM
46 is not suited to (nmigen.FSM is clock-driven, and some aspects of
47 the nested FSMs below are *combinatorial*).
49 * One FSM covers Operand collection and communication address-side
50 with the LD/ST PortInterface. its role ends when "RD_DONE" is asserted
52 * A second FSM activates to cover LD. it activates if op_is_ld is true
54 * A third FSM activates to cover ST. it activates if op_is_st is true
56 * The "overall" (fourth) FSM coordinates the progression and completion
57 of the three other FSMs, firing "WR_RESET" which switches off "busy"
61 https://libre-soc.org/3d_gpu/ld_st_comp_unit.jpg
63 Links including to walk-through videos:
65 * https://libre-soc.org/3d_gpu/architecture/6600scoreboard/
66 * http://libre-soc.org/openpower/isa/fixedload
67 * http://libre-soc.org/openpower/isa/fixedstore
71 * https://bugs.libre-soc.org/show_bug.cgi?id=302
72 * https://bugs.libre-soc.org/show_bug.cgi?id=216
76 * EA - Effective Address
81 from nmigen
.compat
.sim
import run_simulation
82 from nmigen
.cli
import verilog
, rtlil
83 from nmigen
import Module
, Signal
, Mux
, Cat
, Elaboratable
, Array
, Repl
84 from nmigen
.hdl
.rec
import Record
, Layout
86 from nmutil
.latch
import SRLatch
, latchregister
87 from nmutil
.byterev
import byte_reverse
88 from nmutil
.extend
import exts
90 from soc
.experiment
.compalu_multi
import go_record
, CompUnitRecord
91 from soc
.experiment
.l0_cache
import PortInterface
92 from soc
.experiment
.pimem
import LDSTException
93 from soc
.fu
.regspec
import RegSpecAPI
95 from openpower
.decoder
.power_enums
import MicrOp
, Function
, LDSTMode
96 from soc
.fu
.ldst
.ldst_input_record
import CompLDSTOpSubset
97 from openpower
.decoder
.power_decoder2
import Data
98 from openpower
.consts
import MSR
99 from soc
.config
.test
.test_loadstore
import TestMemPspec
102 from nmutil
.util
import Display
105 # TODO: LDSTInputData and LDSTOutputData really should be used
106 # here, to make things more like the other CompUnits. currently,
107 # also, RegSpecAPI is used explicitly here
110 class LDSTCompUnitRecord(CompUnitRecord
):
111 def __init__(self
, rwid
, opsubset
=CompLDSTOpSubset
, name
=None):
112 CompUnitRecord
.__init
__(self
, opsubset
, rwid
,
113 n_src
=3, n_dst
=2, name
=name
)
115 self
.ad
= go_record(1, name
="cu_ad") # address go in, req out
116 self
.st
= go_record(1, name
="cu_st") # store go in, req out
118 self
.exc_o
= LDSTException("exc_o")
120 self
.ld_o
= Signal(reset_less
=True) # operation is a LD
121 self
.st_o
= Signal(reset_less
=True) # operation is a ST
123 # hmm... are these necessary?
124 self
.load_mem_o
= Signal(reset_less
=True) # activate memory LOAD
125 self
.stwd_mem_o
= Signal(reset_less
=True) # activate memory STORE
128 class LDSTCompUnit(RegSpecAPI
, Elaboratable
):
129 """LOAD / STORE Computation Unit
134 * :pi: a PortInterface to the memory subsystem (read-write capable)
135 * :rwid: register width
136 * :awid: address width
140 * :src_i: Source Operands (RA/RB/RC) - managed by rd[0-3] go/req
144 * :data_o: Dest out (LD) - managed by wr[0] go/req
145 * :addr_o: Address out (LD or ST) - managed by wr[1] go/req
146 * :exc_o: Address/Data Exception occurred. LD/ST must terminate
148 TODO: make exc_o a data-type rather than a single-bit signal
154 * :oper_i: operation being carried out (POWER9 decode LD/ST subset)
155 * :issue_i: LD/ST is being "issued".
156 * :shadown_i: Inverted-shadow is being held (stops STORE *and* WRITE)
157 * :go_rd_i: read is being actioned (latches in src regs)
158 * :go_wr_i: write mode (exactly like ALU CompUnit)
159 * :go_ad_i: address is being actioned (triggers actual mem LD)
160 * :go_st_i: store is being actioned (triggers actual mem STORE)
161 * :go_die_i: resets the unit back to "wait for issue"
163 Control Signals (Out)
164 ---------------------
166 * :busy_o: function unit is busy
167 * :rd_rel_o: request src1/src2
168 * :adr_rel_o: request address (from mem)
169 * :sto_rel_o: request store (to mem)
170 * :req_rel_o: request write (result)
171 * :load_mem_o: activate memory LOAD
172 * :stwd_mem_o: activate memory STORE
174 Note: load_mem_o, stwd_mem_o and req_rel_o MUST all be acknowledged
175 in a single cycle and the CompUnit set back to doing another op.
176 This means deasserting go_st_i, go_ad_i or go_wr_i as appropriate
177 depending on whether the operation is a ST or LD.
179 Note: LDSTCompUnit takes care of LE/BE normalisation:
180 * LD data is normalised after receipt from the PortInterface
181 * ST data is normalised *prior* to sending onto the PortInterface
182 TODO: use one module for the byte-reverse as it's quite expensive in gates
185 def __init__(self
, pi
=None, rwid
=64, awid
=48, opsubset
=CompLDSTOpSubset
,
186 debugtest
=False, name
=None):
187 super().__init
__(rwid
)
190 self
.cu
= cu
= LDSTCompUnitRecord(rwid
, opsubset
, name
=name
)
191 self
.debugtest
= debugtest
193 # POWER-compliant LD/ST has index and update: *fixed* number of ports
194 self
.n_src
= n_src
= 3 # RA, RB, RT/RS
195 self
.n_dst
= n_dst
= 2 # RA, RT/RS
197 # set up array of src and dest signals
198 for i
in range(n_src
):
199 j
= i
+ 1 # name numbering to match src1/src2
201 setattr(self
, name
, getattr(cu
, name
))
204 for i
in range(n_dst
):
205 j
= i
+ 1 # name numbering to match dest1/2...
206 name
= "dest%d_o" % j
207 setattr(self
, name
, getattr(cu
, name
))
212 self
.rdmaskn
= cu
.rdmaskn
213 self
.wrmask
= cu
.wrmask
218 # HACK: get data width from dest[0]. this is used across the board
219 # (it really shouldn't be)
220 self
.data_wid
= self
.dest
[0].shape()
222 self
.go_rd_i
= self
.rd
.go_i
# temporary naming
223 self
.go_wr_i
= self
.wr
.go_i
# temporary naming
224 self
.go_ad_i
= self
.ad
.go_i
# temp naming: go address in
225 self
.go_st_i
= self
.st
.go_i
# temp naming: go store in
227 self
.rd_rel_o
= self
.rd
.rel_o
# temporary naming
228 self
.req_rel_o
= self
.wr
.rel_o
# temporary naming
229 self
.adr_rel_o
= self
.ad
.rel_o
# request address (from mem)
230 self
.sto_rel_o
= self
.st
.rel_o
# request store (to mem)
232 self
.issue_i
= cu
.issue_i
233 self
.shadown_i
= cu
.shadown_i
234 self
.go_die_i
= cu
.go_die_i
236 self
.oper_i
= cu
.oper_i
237 self
.src_i
= cu
._src
_i
239 self
.data_o
= Data(self
.data_wid
, name
="o") # Dest1 out: RT
240 self
.addr_o
= Data(self
.data_wid
, name
="ea") # Addr out: Update => RA
241 self
.exc_o
= cu
.exc_o
242 self
.done_o
= cu
.done_o
243 self
.busy_o
= cu
.busy_o
248 self
.load_mem_o
= cu
.load_mem_o
249 self
.stwd_mem_o
= cu
.stwd_mem_o
251 def elaborate(self
, platform
):
257 issue_i
= self
.issue_i
259 #####################
260 # latches for the FSM.
261 m
.submodules
.opc_l
= opc_l
= SRLatch(sync
=False, name
="opc")
262 m
.submodules
.src_l
= src_l
= SRLatch(False, self
.n_src
, name
="src")
263 m
.submodules
.alu_l
= alu_l
= SRLatch(sync
=False, name
="alu")
264 m
.submodules
.adr_l
= adr_l
= SRLatch(sync
=False, name
="adr")
265 m
.submodules
.lod_l
= lod_l
= SRLatch(sync
=False, name
="lod")
266 m
.submodules
.sto_l
= sto_l
= SRLatch(sync
=False, name
="sto")
267 m
.submodules
.wri_l
= wri_l
= SRLatch(sync
=False, name
="wri")
268 m
.submodules
.upd_l
= upd_l
= SRLatch(sync
=False, name
="upd")
269 m
.submodules
.rst_l
= rst_l
= SRLatch(sync
=False, name
="rst")
270 m
.submodules
.lsd_l
= lsd_l
= SRLatch(sync
=False, name
="lsd") # done
276 op_is_ld
= Signal(reset_less
=True)
277 op_is_st
= Signal(reset_less
=True)
279 # ALU/LD data output control
280 alu_valid
= Signal(reset_less
=True) # ALU operands are valid
281 alu_ok
= Signal(reset_less
=True) # ALU out ok (1 clock delay valid)
282 addr_ok
= Signal(reset_less
=True) # addr ok (from PortInterface)
283 ld_ok
= Signal(reset_less
=True) # LD out ok from PortInterface
284 wr_any
= Signal(reset_less
=True) # any write (incl. store)
285 rda_any
= Signal(reset_less
=True) # any read for address ops
286 rd_done
= Signal(reset_less
=True) # all *necessary* operands read
287 wr_reset
= Signal(reset_less
=True) # final reset condition
290 alu_o
= Signal(self
.data_wid
, reset_less
=True)
291 ldd_o
= Signal(self
.data_wid
, reset_less
=True)
293 ##############################
294 # reset conditions for latches
296 # temporaries (also convenient when debugging)
297 reset_o
= Signal(reset_less
=True) # reset opcode
298 reset_w
= Signal(reset_less
=True) # reset write
299 reset_u
= Signal(reset_less
=True) # reset update
300 reset_a
= Signal(reset_less
=True) # reset adr latch
301 reset_i
= Signal(reset_less
=True) # issue|die (use a lot)
302 reset_r
= Signal(self
.n_src
, reset_less
=True) # reset src
303 reset_s
= Signal(reset_less
=True) # reset store
305 comb
+= reset_i
.eq(issue_i | self
.go_die_i
) # various
306 comb
+= reset_o
.eq(self
.done_o | self
.go_die_i
) # opcode reset
307 comb
+= reset_w
.eq(self
.wr
.go_i
[0] | self
.go_die_i
) # write reg 1
308 comb
+= reset_u
.eq(self
.wr
.go_i
[1] | self
.go_die_i
) # update (reg 2)
309 comb
+= reset_s
.eq(self
.go_st_i | self
.go_die_i
) # store reset
310 comb
+= reset_r
.eq(self
.rd
.go_i |
Repl(self
.go_die_i
, self
.n_src
))
311 comb
+= reset_a
.eq(self
.go_ad_i | self
.go_die_i
)
313 p_st_go
= Signal(reset_less
=True)
314 sync
+= p_st_go
.eq(self
.st
.go_i
)
316 # decode bits of operand (latched)
317 oper_r
= CompLDSTOpSubset(name
="oper_r") # Dest register
318 comb
+= op_is_st
.eq(oper_r
.insn_type
== MicrOp
.OP_STORE
) # ST
319 comb
+= op_is_ld
.eq(oper_r
.insn_type
== MicrOp
.OP_LOAD
) # LD
320 comb
+= Display("compldst_multi: op_is_dcbz = %i",
321 (oper_r
.insn_type
== MicrOp
.OP_DCBZ
))
322 op_is_update
= oper_r
.ldst_mode
== LDSTMode
.update
# UPDATE
323 op_is_cix
= oper_r
.ldst_mode
== LDSTMode
.cix
# cache-inhibit
324 comb
+= self
.load_mem_o
.eq(op_is_ld
& self
.go_ad_i
)
325 comb
+= self
.stwd_mem_o
.eq(op_is_st
& self
.go_st_i
)
326 comb
+= self
.ld_o
.eq(op_is_ld
)
327 comb
+= self
.st_o
.eq(op_is_st
)
329 ##########################
330 # FSM implemented through sequence of latches. approximately this:
332 # - src_l[0] : operands
334 # - alu_l : looks after add of src1/2/imm (EA)
335 # - adr_l : waits for add (EA)
336 # - upd_l : waits for adr and Regfile (port 2)
338 # - lod_l : waits for adr (EA) and for LD Data
339 # - wri_l : waits for LD Data and Regfile (port 1)
340 # - st_l : waits for alu and operand2
341 # - rst_l : waits for all FSM paths to converge.
342 # NOTE: use sync to stop combinatorial loops.
344 # opcode latch - inverted so that busy resets to 0
345 # note this MUST be sync so as to avoid a combinatorial loop
346 # between busy_o and issue_i on the reset latch (rst_l)
347 sync
+= opc_l
.s
.eq(issue_i
) # XXX NOTE: INVERTED FROM book!
348 sync
+= opc_l
.r
.eq(reset_o
) # XXX NOTE: INVERTED FROM book!
351 sync
+= src_l
.s
.eq(Repl(issue_i
, self
.n_src
))
352 sync
+= src_l
.r
.eq(reset_r
)
354 # alu latch. use sync-delay between alu_ok and valid to generate pulse
355 comb
+= alu_l
.s
.eq(reset_i
)
356 comb
+= alu_l
.r
.eq(alu_ok
& ~alu_valid
& ~rda_any
)
359 comb
+= adr_l
.s
.eq(reset_i
)
360 sync
+= adr_l
.r
.eq(reset_a
)
363 comb
+= lod_l
.s
.eq(reset_i
)
364 comb
+= lod_l
.r
.eq(ld_ok
)
367 comb
+= wri_l
.s
.eq(issue_i
)
368 sync
+= wri_l
.r
.eq(reset_w |
Repl(wr_reset |
369 (~self
.pi
.busy_o
& op_is_update
),
370 #(self.pi.busy_o & op_is_update),
371 #self.done_o | (self.pi.busy_o & op_is_update),
374 # update-mode operand latch (EA written to reg 2)
375 sync
+= upd_l
.s
.eq(reset_i
)
376 sync
+= upd_l
.r
.eq(reset_u
)
379 comb
+= sto_l
.s
.eq(addr_ok
& op_is_st
)
380 sync
+= sto_l
.r
.eq(reset_s | p_st_go
)
382 # ld/st done. needed to stop LD/ST from activating repeatedly
383 comb
+= lsd_l
.s
.eq(issue_i
)
384 sync
+= lsd_l
.r
.eq(reset_s | p_st_go | ld_ok
)
387 comb
+= rst_l
.s
.eq(addr_ok
) # start when address is ready
388 comb
+= rst_l
.r
.eq(issue_i
)
390 # create a latch/register for the operand
391 with m
.If(self
.issue_i
):
392 sync
+= oper_r
.eq(self
.oper_i
)
393 with m
.If(self
.done_o
):
397 ldd_r
= Signal(self
.data_wid
, reset_less
=True) # Dest register
398 latchregister(m
, ldd_o
, ldd_r
, ld_ok
, name
="ldo_r")
400 # and for each input from the incoming src operands
402 for i
in range(self
.n_src
):
404 src_r
= Signal(self
.data_wid
, name
=name
, reset_less
=True)
405 with m
.If(self
.rd
.go_i
[i
]):
406 sync
+= src_r
.eq(self
.src_i
[i
])
407 with m
.If(self
.issue_i
):
411 # and one for the output from the ADD (for the EA)
412 addr_r
= Signal(self
.data_wid
, reset_less
=True) # Effective Address
413 latchregister(m
, alu_o
, addr_r
, alu_l
.q
, "ea_r")
415 # select either zero or src1 if opcode says so
416 op_is_z
= oper_r
.zero_a
417 src1_or_z
= Signal(self
.data_wid
, reset_less
=True)
418 m
.d
.comb
+= src1_or_z
.eq(Mux(op_is_z
, 0, srl
[0]))
420 # select either immediate or src2 if opcode says so
421 op_is_imm
= oper_r
.imm_data
.ok
422 src2_or_imm
= Signal(self
.data_wid
, reset_less
=True)
423 m
.d
.comb
+= src2_or_imm
.eq(Mux(op_is_imm
, oper_r
.imm_data
.data
, srl
[1]))
425 # now do the ALU addr add: one cycle, and say "ready" (next cycle, too)
426 comb
+= alu_o
.eq(src1_or_z
+ src2_or_imm
) # actual EA
427 m
.d
.sync
+= alu_ok
.eq(alu_valid
) # keep ack in sync with EA
429 ############################
430 # Control Signal calculation
434 comb
+= self
.busy_o
.eq(opc_l
.q
) # | self.pi.busy_o) # busy out
436 # 1st operand read-request only when zero not active
437 # 2nd operand only needed when immediate is not active
438 slg
= Cat(op_is_z
, op_is_imm
)
439 bro
= Repl(self
.busy_o
, self
.n_src
)
440 comb
+= self
.rd
.rel_o
.eq(src_l
.q
& bro
& ~slg
& ~self
.rdmaskn
)
442 # note when the address-related read "go" signals are active
443 comb
+= rda_any
.eq(self
.rd
.go_i
[0] | self
.rd
.go_i
[1])
445 # alu input valid when 1st and 2nd ops done (or imm not active)
446 comb
+= alu_valid
.eq(busy_o
& ~
(self
.rd
.rel_o
[0] | self
.rd
.rel_o
[1]))
448 # 3rd operand only needed when operation is a store
449 comb
+= self
.rd
.rel_o
[2].eq(src_l
.q
[2] & busy_o
& op_is_st
)
451 # all reads done when alu is valid and 3rd operand needed
452 comb
+= rd_done
.eq(alu_valid
& ~self
.rd
.rel_o
[2])
454 # address release only if addr ready, but Port must be idle
455 comb
+= self
.adr_rel_o
.eq(alu_valid
& adr_l
.q
& busy_o
)
457 # the write/store (etc) all must be cancelled if an exception occurs
458 cancel
= Signal(reset_less
=True)
459 comb
+= cancel
.eq(self
.exc_o
.happened | self
.shadown_i
)
461 # store release when st ready *and* all operands read (and no shadow)
462 comb
+= self
.st
.rel_o
.eq(sto_l
.q
& busy_o
& rd_done
& op_is_st
&
465 # request write of LD result. waits until shadow is dropped.
466 comb
+= self
.wr
.rel_o
[0].eq(rd_done
& wri_l
.q
& busy_o
& lod_l
.qn
&
469 # request write of EA result only in update mode
470 comb
+= self
.wr
.rel_o
[1].eq(upd_l
.q
& busy_o
& op_is_update
&
473 # provide "done" signal: select req_rel for non-LD/ST, adr_rel for LD/ST
474 comb
+= wr_any
.eq(self
.st
.go_i | p_st_go |
475 self
.wr
.go_i
[0] | self
.wr
.go_i
[1])
476 comb
+= wr_reset
.eq(rst_l
.q
& busy_o
& cancel
&
477 ~
(self
.st
.rel_o | self
.wr
.rel_o
[0] |
479 (lod_l
.qn | op_is_st
)
481 comb
+= self
.done_o
.eq(wr_reset
& (~self
.pi
.busy_o | op_is_ld
))
483 ######################
484 # Data/Address outputs
486 # put the LD-output register directly onto the output bus on a go_write
487 comb
+= self
.data_o
.data
.eq(self
.dest
[0])
488 with m
.If(self
.wr
.go_i
[0]):
489 comb
+= self
.dest
[0].eq(ldd_r
)
491 # "update" mode, put address out on 2nd go-write
492 comb
+= self
.addr_o
.data
.eq(self
.dest
[1])
493 with m
.If(op_is_update
& self
.wr
.go_i
[1]):
494 comb
+= self
.dest
[1].eq(addr_r
)
496 # need to look like MultiCompUnit: put wrmask out.
497 # XXX may need to make this enable only when write active
498 comb
+= self
.wrmask
.eq(bro
& Cat(op_is_ld
, op_is_update
))
500 ###########################
501 # PortInterface connections
504 # connect to LD/ST PortInterface.
505 comb
+= pi
.is_ld_i
.eq(op_is_ld
& busy_o
) # decoded-LD
506 comb
+= pi
.is_st_i
.eq(op_is_st
& busy_o
) # decoded-ST
507 comb
+= pi
.data_len
.eq(oper_r
.data_len
) # data_len
508 # address: use sync to avoid long latency
509 sync
+= pi
.addr
.data
.eq(addr_r
) # EA from adder
510 sync
+= pi
.addr
.ok
.eq(alu_ok
& lsd_l
.q
) # "do address stuff" (once)
511 comb
+= self
.exc_o
.eq(pi
.exc_o
) # exception occurred
512 comb
+= addr_ok
.eq(self
.pi
.addr_ok_o
) # no exc, address fine
513 # connect MSR.PR for priv/virt operation
514 comb
+= pi
.msr_pr
.eq(oper_r
.msr
[MSR
.PR
])
517 revnorev
= Signal(64, reset_less
=True)
518 with m
.If(oper_r
.byte_reverse
):
519 # byte-reverse the data based on ld/st width (turn it to LE)
520 data_len
= oper_r
.data_len
521 lddata_r
= byte_reverse(m
, 'lddata_r', pi
.ld
.data
, data_len
)
522 comb
+= revnorev
.eq(lddata_r
) # put reversed- data out
524 comb
+= revnorev
.eq(pi
.ld
.data
) # put data out, straight (as BE)
526 # then check sign-extend
527 with m
.If(oper_r
.sign_extend
):
528 # okok really should "if data_len == 4" and so on here
529 with m
.If(oper_r
.data_len
== 2):
530 comb
+= ldd_o
.eq(exts(revnorev
, 16, 64)) # sign-extend hword
532 comb
+= ldd_o
.eq(exts(revnorev
, 32, 64)) # sign-extend dword
534 comb
+= ldd_o
.eq(revnorev
)
536 # ld - ld gets latched in via lod_l
537 comb
+= ld_ok
.eq(pi
.ld
.ok
) # ld.ok *closes* (freezes) ld data
540 op3
= srl
[2] # 3rd operand latch
541 with m
.If(oper_r
.byte_reverse
):
542 # byte-reverse the data based on width
543 data_len
= oper_r
.data_len
544 stdata_r
= byte_reverse(m
, 'stdata_r', op3
, data_len
)
545 comb
+= pi
.st
.data
.eq(stdata_r
)
547 comb
+= pi
.st
.data
.eq(op3
)
548 # store - data goes in based on go_st
549 comb
+= pi
.st
.ok
.eq(self
.st
.go_i
) # go store signals st data valid
553 def get_out(self
, i
):
554 """make LDSTCompUnit look like RegSpecALUAPI. these correspond
555 to LDSTOutputData o and o1 respectively.
558 return self
.data_o
# LDSTOutputData.regspec o
560 return self
.addr_o
# LDSTOutputData.regspec o1
561 # return self.dest[i]
563 def get_fu_out(self
, i
):
564 return self
.get_out(i
)
574 yield from self
.oper_i
.ports()
575 yield from self
.src_i
581 yield from self
.data_o
.ports()
582 yield from self
.addr_o
.ports()
583 yield self
.load_mem_o
584 yield self
.stwd_mem_o
590 def wait_for(sig
, wait
=True, test1st
=False):
592 print("wait for", sig
, v
, wait
, test1st
)
593 if test1st
and bool(v
) == wait
:
598 #print("...wait for", sig, v)
603 def store(dut
, src1
, src2
, src3
, imm
, imm_ok
=True, update
=False,
605 print("ST", src1
, src2
, src3
, imm
, imm_ok
, update
)
606 yield dut
.oper_i
.insn_type
.eq(MicrOp
.OP_STORE
)
607 yield dut
.oper_i
.data_len
.eq(2) # half-word
608 yield dut
.oper_i
.byte_reverse
.eq(byterev
)
609 yield dut
.src1_i
.eq(src1
)
610 yield dut
.src2_i
.eq(src2
)
611 yield dut
.src3_i
.eq(src3
)
612 yield dut
.oper_i
.imm_data
.imm
.eq(imm
)
613 yield dut
.oper_i
.imm_data
.ok
.eq(imm_ok
)
614 yield dut
.oper_i
.update
.eq(update
)
615 yield dut
.issue_i
.eq(1)
617 yield dut
.issue_i
.eq(0)
623 # wait for all active rel signals to come up
625 rel
= yield dut
.rd
.rel_o
626 if rel
== active_rel
:
629 yield dut
.rd
.go
.eq(active_rel
)
631 yield dut
.rd
.go
.eq(0)
633 yield from wait_for(dut
.adr_rel_o
, False, test1st
=True)
634 # yield from wait_for(dut.adr_rel_o)
635 # yield dut.ad.go.eq(1)
637 # yield dut.ad.go.eq(0)
640 yield from wait_for(dut
.wr
.rel_o
[1])
641 yield dut
.wr
.go
.eq(0b10)
643 addr
= yield dut
.addr_o
645 yield dut
.wr
.go
.eq(0)
649 yield from wait_for(dut
.sto_rel_o
)
650 yield dut
.go_st_i
.eq(1)
652 yield dut
.go_st_i
.eq(0)
653 yield from wait_for(dut
.busy_o
, False)
654 # wait_for(dut.stwd_mem_o)
659 def load(dut
, src1
, src2
, imm
, imm_ok
=True, update
=False, zero_a
=False,
661 print("LD", src1
, src2
, imm
, imm_ok
, update
)
662 yield dut
.oper_i
.insn_type
.eq(MicrOp
.OP_LOAD
)
663 yield dut
.oper_i
.data_len
.eq(2) # half-word
664 yield dut
.oper_i
.byte_reverse
.eq(byterev
)
665 yield dut
.src1_i
.eq(src1
)
666 yield dut
.src2_i
.eq(src2
)
667 yield dut
.oper_i
.zero_a
.eq(zero_a
)
668 yield dut
.oper_i
.imm_data
.imm
.eq(imm
)
669 yield dut
.oper_i
.imm_data
.ok
.eq(imm_ok
)
670 yield dut
.issue_i
.eq(1)
672 yield dut
.issue_i
.eq(0)
675 # set up read-operand flags
677 if not imm_ok
: # no immediate means RB register needs to be read
679 if not zero_a
: # no zero-a means RA needs to be read
682 # wait for the operands (RA, RB, or both)
684 yield dut
.rd
.go
.eq(rd
)
685 yield from wait_for(dut
.rd
.rel_o
)
686 yield dut
.rd
.go
.eq(0)
688 yield from wait_for(dut
.adr_rel_o
, False, test1st
=True)
689 # yield dut.ad.go.eq(1)
691 # yield dut.ad.go.eq(0)
694 yield from wait_for(dut
.wr
.rel_o
[1])
695 yield dut
.wr
.go
.eq(0b10)
697 addr
= yield dut
.addr_o
699 yield dut
.wr
.go
.eq(0)
703 yield from wait_for(dut
.wr
.rel_o
[0], test1st
=True)
704 yield dut
.wr
.go
.eq(1)
706 data
= yield dut
.data_o
708 yield dut
.wr
.go
.eq(0)
709 yield from wait_for(dut
.busy_o
)
711 # wait_for(dut.stwd_mem_o)
720 # two STs (different addresses)
721 yield from store(dut
, 4, 0, 3, 2) # ST reg4 into addr rfile[reg3]+2
722 yield from store(dut
, 2, 0, 9, 2) # ST reg4 into addr rfile[reg9]+2
724 # two LDs (deliberately LD from the 1st address then 2nd)
725 data
, addr
= yield from load(dut
, 4, 0, 2)
726 assert data
== 0x0003, "returned %x" % data
727 data
, addr
= yield from load(dut
, 2, 0, 2)
728 assert data
== 0x0009, "returned %x" % data
732 yield from store(dut
, 9, 5, 3, 0, imm_ok
=False)
733 data
, addr
= yield from load(dut
, 9, 5, 0, imm_ok
=False)
734 assert data
== 0x0003, "returned %x" % data
736 # update-immediate version
737 addr
= yield from store(dut
, 9, 6, 3, 2, update
=True)
738 assert addr
== 0x000b, "returned %x" % addr
740 # update-indexed version
741 data
, addr
= yield from load(dut
, 9, 5, 0, imm_ok
=False, update
=True)
742 assert data
== 0x0003, "returned %x" % data
743 assert addr
== 0x000e, "returned %x" % addr
745 # immediate *and* zero version
746 data
, addr
= yield from load(dut
, 1, 4, 8, imm_ok
=True, zero_a
=True)
747 assert data
== 0x0008, "returned %x" % data
750 class TestLDSTCompUnit(LDSTCompUnit
):
752 def __init__(self
, rwid
, pspec
):
753 from soc
.experiment
.l0_cache
import TstL0CacheBuffer
754 self
.l0
= l0
= TstL0CacheBuffer(pspec
)
755 pi
= l0
.l0
.dports
[0].pi
756 LDSTCompUnit
.__init
__(self
, pi
, rwid
, 4)
758 def elaborate(self
, platform
):
759 m
= LDSTCompUnit
.elaborate(self
, platform
)
760 m
.submodules
.l0
= self
.l0
761 m
.d
.comb
+= self
.ad
.go
.eq(self
.ad
.rel
) # link addr-go direct to rel
765 def test_scoreboard():
768 pspec
= TestMemPspec(ldst_ifacetype
='bare_wb',
769 imem_ifacetype
='bare_wb',
775 dut
= TestLDSTCompUnit(16,pspec
)
776 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
777 with
open("test_ldst_comp.il", "w") as f
:
780 run_simulation(dut
, ldst_sim(dut
), vcd_name
='test_ldst_comp.vcd')
783 class TestLDSTCompUnitRegSpec(LDSTCompUnit
):
785 def __init__(self
, pspec
):
786 from soc
.experiment
.l0_cache
import TstL0CacheBuffer
787 from soc
.fu
.ldst
.pipe_data
import LDSTPipeSpec
788 regspec
= LDSTPipeSpec
.regspec
789 self
.l0
= l0
= TstL0CacheBuffer(pspec
)
790 pi
= l0
.l0
.dports
[0].pi
791 LDSTCompUnit
.__init
__(self
, pi
, regspec
, 4)
793 def elaborate(self
, platform
):
794 m
= LDSTCompUnit
.elaborate(self
, platform
)
795 m
.submodules
.l0
= self
.l0
796 m
.d
.comb
+= self
.ad
.go
.eq(self
.ad
.rel
) # link addr-go direct to rel
800 def test_scoreboard_regspec():
803 pspec
= TestMemPspec(ldst_ifacetype
='bare_wb',
804 imem_ifacetype
='bare_wb',
810 dut
= TestLDSTCompUnitRegSpec(pspec
)
811 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
812 with
open("test_ldst_comp.il", "w") as f
:
815 run_simulation(dut
, ldst_sim(dut
), vcd_name
='test_ldst_regspec.vcd')
818 if __name__
== '__main__':
819 test_scoreboard_regspec()