Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 19:47:27 +0000 (19:47 +0000)]
stack of changes to MultiCompUnit to speed it up
a lot of comb settings were changed to sync as a hack-experiment
when TestIssuer (ls180) was being developed, to try to speed up ECP5
timing. those need reverting to reduce latency
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 16:25:10 +0000 (16:25 +0000)]
experimenting with option to shorten MultiCompUnit delays
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 16:24:50 +0000 (16:24 +0000)]
create single-stage ALU pipeline, shorten latency on in-order core
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 15:59:27 +0000 (15:59 +0000)]
allow MultiCompUnit to set read and write latches to run on sync or comb
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 13:41:21 +0000 (13:41 +0000)]
FunctionUnitBaseMulti which derives from ReservationStations2
*requires* the muxid to be set for each RS. these can be set
combinatorially no problem, to indicate to ReservationStations2 where
to route the data after it pops out the end of the "real" ALU
without these muxids being set everything is royally screwed
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 13:01:01 +0000 (13:01 +0000)]
better name for read latch in core.py
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 13:00:41 +0000 (13:00 +0000)]
use m.submodules[name] instead of getattr
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 12:20:23 +0000 (12:20 +0000)]
remove redundant / mis-named variable in core
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 12:19:33 +0000 (12:19 +0000)]
code-comments
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 12:16:18 +0000 (12:16 +0000)]
remove unneeded data structure in core
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 12:12:27 +0000 (12:12 +0000)]
whoops treereduce on write-vector set/clr error
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 12:10:48 +0000 (12:10 +0000)]
more code-cleanup
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 12:07:45 +0000 (12:07 +0000)]
use new regspec_decode and fu.get_iospec functions
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 12:04:03 +0000 (12:04 +0000)]
core tidyup
Luke Kenneth Casson Leighton [Wed, 1 Dec 2021 11:53:27 +0000 (11:53 +0000)]
add Regspecs get_io_spec function
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 18:29:41 +0000 (18:29 +0000)]
start allocating more FUs (more ReservationStations)
Tobias Platen [Tue, 30 Nov 2021 17:52:10 +0000 (18:52 +0100)]
random loadstore1 test: readback written data
Tobias Platen [Tue, 30 Nov 2021 17:29:58 +0000 (18:29 +0100)]
reenable dcbz test case
Tobias Platen [Tue, 30 Nov 2021 17:29:27 +0000 (18:29 +0100)]
return correct data from microwatt
Tobias Platen [Tue, 30 Nov 2021 17:22:16 +0000 (18:22 +0100)]
cleanup test_loadstore1.py
Tobias Platen [Tue, 30 Nov 2021 16:56:45 +0000 (17:56 +0100)]
loadstore: add done_delay
Tobias Platen [Tue, 30 Nov 2021 16:52:40 +0000 (17:52 +0100)]
Merge branch 'master' of ssh://git.libre-riscv.org:922/soc
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 16:25:45 +0000 (16:25 +0000)]
add LogicalTestCases back in to test_core.py (pass)
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 16:23:43 +0000 (16:23 +0000)]
let PowerDecode2 decide which operand class to use in test_core.py
this gets it to pick *all* fields
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 16:12:31 +0000 (16:12 +0000)]
use latched readflag (recspec_decode_read "ok") instead of global one
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 16:11:58 +0000 (16:11 +0000)]
tidyup on read-flag latches
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 14:05:38 +0000 (14:05 +0000)]
fix read-decode information by latching not just the read reg number
but *whether it is valid or not*.
not only the register *number* from regspec_decode_read() needs to
be captured but its ok (valid) flag as well
this error only showed up when a mul (2-operand) was followed by addi
(1-operand). the mul had its 2nd operand dropped on the floor because
the *addi* only has one operand.
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 13:21:29 +0000 (13:21 +0000)]
fix write-after-write hazard checking
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 12:11:45 +0000 (12:11 +0000)]
allow busy to settle before checking state in test_core.py
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 12:03:29 +0000 (12:03 +0000)]
only check regs right at the end in test_core.py overlap mode
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 11:49:33 +0000 (11:49 +0000)]
move sim call before core run in test_core.py
to give PC a chance to settle in regfile
Luke Kenneth Casson Leighton [Tue, 30 Nov 2021 11:13:24 +0000 (11:13 +0000)]
getting formerly unused test_core.py operational
Luke Kenneth Casson Leighton [Mon, 29 Nov 2021 16:41:00 +0000 (16:41 +0000)]
whoops missed make_hazard_vec test
Luke Kenneth Casson Leighton [Mon, 29 Nov 2021 15:50:26 +0000 (15:50 +0000)]
whoops do the set/get of the write-vector at a single access point
for example: INT o, INT o2 is *two* write ports. therefore,
previously, there was a comb += wvint bitvector for o, which was
then OVERWRITTEN by a second comb += wvint bitvector for o2
instead, return the wvset/wvclr requests to connect_wrport, which
collates by regfile name and performs an additional or-tree-reduce
on the requests to set/get
Luke Kenneth Casson Leighton [Mon, 29 Nov 2021 13:43:36 +0000 (13:43 +0000)]
always set fwd_bus_mode=False on regfiles
Luke Kenneth Casson Leighton [Mon, 29 Nov 2021 13:19:45 +0000 (13:19 +0000)]
add MMU and SPR to list of FUs that must report "busy" i.e. block
Luke Kenneth Casson Leighton [Mon, 29 Nov 2021 10:26:35 +0000 (10:26 +0000)]
disallow overlap in core on LDST, Branch, and Trap.
these are all the ones that could cause exceptions or mess with ordering
Luke Kenneth Casson Leighton [Mon, 29 Nov 2021 10:22:09 +0000 (10:22 +0000)]
use dict style not setattr on submodules
Tobias Platen [Sun, 28 Nov 2021 13:54:32 +0000 (14:54 +0100)]
update loadstore1 testcase
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 14:32:21 +0000 (14:32 +0000)]
code-comments
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 14:11:23 +0000 (14:11 +0000)]
fix instructions of the type "read-reg-is-same-as-write"
this then caused write-after-write detection to go into a combinatorial
loop, which was fixed by not checking fu.issue_i but instead using
fu_enable (fu_bitdict[FU])
Tobias Platen [Sat, 27 Nov 2021 12:54:22 +0000 (13:54 +0100)]
loadstore testcase: read at random addresses
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 12:51:52 +0000 (12:51 +0000)]
FU-Regs matrix tidyup and comments
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 12:33:02 +0000 (12:33 +0000)]
minor tidyup on FU-Regs Matrix
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 12:20:33 +0000 (12:20 +0000)]
update FURegDepMatrix to multi-dest
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 12:19:57 +0000 (12:19 +0000)]
update naming on Reg_Rsv signals
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 11:54:37 +0000 (11:54 +0000)]
add copyright and attribution notices
update Reg_Rsv to multi-dest
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 11:44:07 +0000 (11:44 +0000)]
update FU_RW_Pending vectors to multi-dest
update Copyright and Attribution notices
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 11:19:57 +0000 (11:19 +0000)]
convert DependencyRow to multiple destination latches
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 10:40:47 +0000 (10:40 +0000)]
add copyright and attribution notices to dependence_cell.py
Luke Kenneth Casson Leighton [Sat, 27 Nov 2021 10:36:30 +0000 (10:36 +0000)]
update license and attribution in fu_reg_matrix.py
Luke Kenneth Casson Leighton [Fri, 26 Nov 2021 16:45:23 +0000 (16:45 +0000)]
convert score6600_multi over to using RegSpecs (in a fake way)
which by slow code-morphing the entire score6600_multi code can be
converted to be like core.py
Luke Kenneth Casson Leighton [Fri, 26 Nov 2021 15:23:40 +0000 (15:23 +0000)]
early use of Array unnecessarily (all uses are static references
not dynamic indexing: no pmux created or needed)
Luke Kenneth Casson Leighton [Fri, 26 Nov 2021 15:19:19 +0000 (15:19 +0000)]
early use of Array unnecessarily (all uses are static references
not dynamic indexing: no pmux created or needed)
Luke Kenneth Casson Leighton [Thu, 25 Nov 2021 21:41:04 +0000 (21:41 +0000)]
get score6600_multi operational again
main thing was a "delay" on go_read_i due to regfile read results being
one clock cycle behind.
Luke Kenneth Casson Leighton [Thu, 25 Nov 2021 21:38:54 +0000 (21:38 +0000)]
add debug prints in old simulator
Luke Kenneth Casson Leighton [Thu, 25 Nov 2021 21:38:42 +0000 (21:38 +0000)]
add debug prints in old simulator
Tobias Platen [Thu, 25 Nov 2021 17:54:21 +0000 (18:54 +0100)]
seperate invalid test case from other test cases
Tobias Platen [Thu, 25 Nov 2021 16:53:42 +0000 (17:53 +0100)]
mmu: add debug output
Tobias Platen [Thu, 25 Nov 2021 16:07:05 +0000 (17:07 +0100)]
add testcase for invalid pagetable
Tobias Platen [Thu, 25 Nov 2021 16:05:25 +0000 (17:05 +0100)]
pimem: reset on exception
Tobias Platen [Thu, 25 Nov 2021 15:49:06 +0000 (16:49 +0100)]
remove unuses dsisr signal
Tobias Platen [Thu, 25 Nov 2021 15:45:10 +0000 (16:45 +0100)]
reset state to idle on exception
Luke Kenneth Casson Leighton [Thu, 25 Nov 2021 09:43:14 +0000 (09:43 +0000)]
more sorting scoremulti
Luke Kenneth Casson Leighton [Thu, 25 Nov 2021 09:28:35 +0000 (09:28 +0000)]
more sorting out scoremulti
Luke Kenneth Casson Leighton [Thu, 25 Nov 2021 09:16:54 +0000 (09:16 +0000)]
add test pspec for scoremulti to work
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 22:46:04 +0000 (22:46 +0000)]
convert hazard bitvectors to Reset-Priority SRLatch from nmutil
the read/write port were seriously getting in the way, by way of the
operand forwarding etc. etc. need instead to get either the immediate-latched
value (combinatorial) or the sync-delayed one, depending on circumstances.
Tobias Platen [Wed, 24 Nov 2021 18:50:05 +0000 (19:50 +0100)]
fix exception handling in pi_ld
Tobias Platen [Wed, 24 Nov 2021 18:30:44 +0000 (19:30 +0100)]
Merge branch 'master' of ssh://git.libre-riscv.org:922/soc
Tobias Platen [Wed, 24 Nov 2021 18:30:25 +0000 (19:30 +0100)]
improve debug output in mmu.py
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 17:48:32 +0000 (17:48 +0000)]
fix write-after-write hazard detection
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 17:44:36 +0000 (17:44 +0000)]
when allow_overlap enabled do a manual wait until all FUs have
entirely finished befor going on to read the regfile state
this really should be done with the DMI interface, which needs updating
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 16:24:03 +0000 (16:24 +0000)]
code-comments
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 16:06:26 +0000 (16:06 +0000)]
add write-after-write hazard detection
stall if a WaW hazard is detected, continuing to assert the main decoder
so that the relevant bits of the write-vector can be read. these
bits are read *one clock behind* which is perfectly fine
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 15:18:08 +0000 (15:18 +0000)]
add 2nd hazard bitvector port for write-after-write
hazard detection
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 12:31:52 +0000 (12:31 +0000)]
whoops merged the two write-ports for RT and RA-with-update
Luke Kenneth Casson Leighton [Wed, 24 Nov 2021 11:55:14 +0000 (11:55 +0000)]
disable hazard vectors when overlap is not requested in core.py
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 22:45:03 +0000 (22:45 +0000)]
update submodules
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 22:44:55 +0000 (22:44 +0000)]
more comments
Tobias Platen [Tue, 23 Nov 2021 18:12:10 +0000 (19:12 +0100)]
pimem changes for st exception handling
Tobias Platen [Tue, 23 Nov 2021 17:42:22 +0000 (18:42 +0100)]
fix test_loadstore1.py
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 15:53:35 +0000 (15:53 +0000)]
add FU write-after-write hazard detection Signal (dummy so far)
this has to be set up in the connect_write section but then
*used* in connect_rdport in order to stop reading from happening
strictly speaking it should actually stop issue, but we already
established that that way lies madness (combinatorial loops).
this may turn out to be a bit more complex than initially planned
due to some instructions potentially not having read at all
(only write)
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 15:21:15 +0000 (15:21 +0000)]
add code-comments, link to in-order core
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 15:03:09 +0000 (15:03 +0000)]
more use of namedtuples in core.py for clarity
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 12:21:36 +0000 (12:21 +0000)]
start some use of namedtuples in core.py
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 12:11:49 +0000 (12:11 +0000)]
use some namedtuples to make things clearer in core.py
Luke Kenneth Casson Leighton [Tue, 23 Nov 2021 12:03:19 +0000 (12:03 +0000)]
use fascinating trick of defaultdict-of-defaultdicts
Tobias Platen [Mon, 22 Nov 2021 18:35:53 +0000 (19:35 +0100)]
add store testcase
Tobias Platen [Mon, 22 Nov 2021 17:57:54 +0000 (18:57 +0100)]
fix fast exception handling for pi_st
Luke Kenneth Casson Leighton [Mon, 22 Nov 2021 14:51:02 +0000 (14:51 +0000)]
make FetchFSM take PC as an input in its ispec
Luke Kenneth Casson Leighton [Mon, 22 Nov 2021 14:41:25 +0000 (14:41 +0000)]
local variable rename in FetchFSM
Luke Kenneth Casson Leighton [Mon, 22 Nov 2021 14:37:37 +0000 (14:37 +0000)]
split out FetchFSM into separate module
there are DriverConflicts which result in the FetchFSM being merged
into TestIssuerInternal but that is actually ok for now
Luke Kenneth Casson Leighton [Mon, 22 Nov 2021 11:59:19 +0000 (11:59 +0000)]
whoops accidentally committed commented-out test for overlap
in test_runner.py
Luke Kenneth Casson Leighton [Sun, 21 Nov 2021 22:57:35 +0000 (22:57 +0000)]
reset execute back to ISSUE_START if at INSN_WAIT and
stopping was requested.
Luke Kenneth Casson Leighton [Sun, 21 Nov 2021 22:44:42 +0000 (22:44 +0000)]
restrict (refine) hazard selection to the one being picked for this port
Luke Kenneth Casson Leighton [Sun, 21 Nov 2021 22:21:52 +0000 (22:21 +0000)]
block picker hazard on input to PriorityPicker rather than output
reason: another register could be picked if the input is blocked
but if the output is masked out, the picker will sit there with the
highest priority port stopping all others
Luke Kenneth Casson Leighton [Sun, 21 Nov 2021 21:12:14 +0000 (21:12 +0000)]
parse test_issuer args allow option "allow-overlap" to be optional
Luke Kenneth Casson Leighton [Sun, 21 Nov 2021 21:10:47 +0000 (21:10 +0000)]
complex. TestRunner now does not work properly unless recognising
that first, instructions can complete out-of-order from which they were
issued, and second, that just because the Function Unit has said it
accepted the instruction does not mean it has actually completed.
TestRunner therefore needed quite a bit of sorting out
Luke Kenneth Casson Leighton [Sun, 21 Nov 2021 21:09:09 +0000 (21:09 +0000)]
fixed issue with hazard dependencies, read will nott
take place until write dependencies are clear. currently extremely
draconian "ban" on entire FU progressing, actually probably all FUs
progressing, until write-hazards are clear. can sort out later
Tobias Platen [Sun, 21 Nov 2021 15:37:11 +0000 (16:37 +0100)]
add testcase for fast exceptions on store