Luke Kenneth Casson Leighton [Thu, 18 Nov 2021 23:19:32 +0000 (23:19 +0000)]
remove combinatorial loop in core instruction conflict detection
Luke Kenneth Casson Leighton [Thu, 18 Nov 2021 22:52:34 +0000 (22:52 +0000)]
experimenting with overlapping instructions, bit of a mess
Luke Kenneth Casson Leighton [Thu, 18 Nov 2021 22:08:00 +0000 (22:08 +0000)]
set up core processing FSM, which captures data if FU is not ready
Luke Kenneth Casson Leighton [Thu, 18 Nov 2021 21:40:48 +0000 (21:40 +0000)]
set up a temporary copy of CoreInput
(and fix CoreInput.eq, sigh)
Luke Kenneth Casson Leighton [Thu, 18 Nov 2021 21:24:19 +0000 (21:24 +0000)]
experiment allowing overlap (activated with --allow-overlap) in TestIssuer
Luke Kenneth Casson Leighton [Thu, 18 Nov 2021 13:25:29 +0000 (13:25 +0000)]
remove unneeded import
Jacob Lifshay [Wed, 17 Nov 2021 20:49:41 +0000 (12:49 -0800)]
start adding bitmanip FU
Tobias Platen [Wed, 17 Nov 2021 18:34:36 +0000 (19:34 +0100)]
PortInterfaceBase: fix fast exception handling
Tobias Platen [Wed, 17 Nov 2021 18:04:20 +0000 (19:04 +0100)]
whitespace
Tobias Platen [Wed, 17 Nov 2021 18:03:23 +0000 (19:03 +0100)]
Merge branch 'master' of ssh://git.libre-riscv.org:922/soc
Tobias Platen [Wed, 17 Nov 2021 18:02:38 +0000 (19:02 +0100)]
fix mistake in test_pi2ls.py
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 18:01:27 +0000 (18:01 +0000)]
reading of regfile bitvector added, which activates on a per-FU basis
at the regfile read port
this is somewhat complete overkill because strictly speaking the
read should be done at issue time. fortunately, merging of lots of ORs
results in the exact same thing, just distributed
horribly inefficient though
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 17:42:05 +0000 (17:42 +0000)]
core hazard bitvector regfiles need to be readable
immediately (combinatorial) not via sync. allow synced option to
pass through from VirtualRegPort to RegFileArray
Tobias Platen [Wed, 17 Nov 2021 17:53:54 +0000 (18:53 +0100)]
fixed busy waiting in pi_st
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 16:23:45 +0000 (16:23 +0000)]
add option to test_issuer.py to allow for overlapping issue of
instructions. this is for Core hazard detection prior to moving to
an in-order core
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 16:15:37 +0000 (16:15 +0000)]
add ability to run hazard instruction for test purposes
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 14:44:40 +0000 (14:44 +0000)]
detect the case in Core bitvector when the Function Unit says:
"actually, although you said i *could* write to this regfile
(and therefore have been reserving a write hazard protection for me)
actually i'm not going to write to it."
this situation occurs when:
* at issue time a wrflag (from PowerDecoder decode_regspec_write) was HI
* at ALU output time (alu.n.o_ready) the FU dest.data.ok flag is LOW
under these unusual but perfectly valid circumstances the write hazard
bitvector MUST still be cleared
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 14:40:56 +0000 (14:40 +0000)]
add probe of whether CompUnit ALU is done or not.
this is used in core to detect whether the optional situation of
an ALU *maybe* writing to a regfile occurs or not
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 14:10:50 +0000 (14:10 +0000)]
missing optional check on make_hazard_vecs
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 14:08:50 +0000 (14:08 +0000)]
move core hazard set/clear to separate function, for clarity
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 13:40:08 +0000 (13:40 +0000)]
whoops context-indentation by mistake (no harm done, just odd)
Luke Kenneth Casson Leighton [Wed, 17 Nov 2021 13:34:52 +0000 (13:34 +0000)]
add a FetchOutput pipeline data structure
Luke Kenneth Casson Leighton [Tue, 16 Nov 2021 19:12:51 +0000 (19:12 +0000)]
print out regfile unary status, bit of name-cleanup
Luke Kenneth Casson Leighton [Tue, 16 Nov 2021 18:54:21 +0000 (18:54 +0000)]
use a virtual regfile port for the hazard bitvectors
this allows a full width of enables and full width of bits
(one per reg being written to)
Tobias Platen [Tue, 16 Nov 2021 18:22:53 +0000 (19:22 +0100)]
pi_ld busy waiting fix
Tobias Platen [Tue, 16 Nov 2021 17:24:59 +0000 (18:24 +0100)]
loadstore1 now reports exception reason
Luke Kenneth Casson Leighton [Tue, 16 Nov 2021 16:29:30 +0000 (16:29 +0000)]
create set/get ports for bitvectors
Luke Kenneth Casson Leighton [Tue, 16 Nov 2021 14:03:28 +0000 (14:03 +0000)]
capture write port (wrflag) in byregfiles_spec for use in
bitvector setting at issue time
Luke Kenneth Casson Leighton [Tue, 16 Nov 2021 13:36:26 +0000 (13:36 +0000)]
rename regports for bitvectors so that
* read regfile can SET the bitvector
* write regfile can CLEAR the bitvector
Luke Kenneth Casson Leighton [Tue, 16 Nov 2021 10:29:37 +0000 (10:29 +0000)]
starting to get write-clear of hazard vectors operating
Luke Kenneth Casson Leighton [Tue, 16 Nov 2021 10:28:46 +0000 (10:28 +0000)]
whoops, hazard vectors were depth 1 width N
they need to be regwidth 1 (1 bit wide) depth N (one per register)
Tobias Platen [Mon, 15 Nov 2021 19:29:13 +0000 (20:29 +0100)]
report dar on exception + test case
Tobias Platen [Mon, 15 Nov 2021 18:48:05 +0000 (19:48 +0100)]
add test_loadstore1.py
Luke Kenneth Casson Leighton [Sat, 13 Nov 2021 15:11:00 +0000 (15:11 +0000)]
add quick instructions on how to run pinouts.py to get some debug info
Luke Kenneth Casson Leighton [Sat, 13 Nov 2021 15:08:43 +0000 (15:08 +0000)]
update submodule to make ngi pointer router pinouts
Luke Kenneth Casson Leighton [Sat, 13 Nov 2021 15:07:23 +0000 (15:07 +0000)]
add new get_pinspec_resources function which creates nmigen
Resource/Subsignal/Pins suite from the JSON files generated by pinmux
Luke Kenneth Casson Leighton [Sat, 13 Nov 2021 14:42:05 +0000 (14:42 +0000)]
code-comment for get_pinspecs()
Luke Kenneth Casson Leighton [Sat, 13 Nov 2021 14:41:51 +0000 (14:41 +0000)]
start adding hazard vector setting in core (unfinished)
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 16:14:58 +0000 (16:14 +0000)]
debug prints
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 16:11:48 +0000 (16:11 +0000)]
fix regfile port names for "fast" port access (regreduce=False)
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 16:10:07 +0000 (16:10 +0000)]
TODO, implement is_dcbz
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 16:00:22 +0000 (16:00 +0000)]
code-comments
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 15:56:05 +0000 (15:56 +0000)]
split out core input/output into separate file core_data.py
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 15:51:39 +0000 (15:51 +0000)]
enable hazard vecs in core
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 14:56:09 +0000 (14:56 +0000)]
add exact same number - and name - bitvector ports to regfiles
for hazard purposes, easier to just have the exact same names
g
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 14:21:49 +0000 (14:21 +0000)]
code-morph regfile port specs to a dictionary format rather than hardcoded
this allows for Hazard Bit-vector regfiles to be created with exactly
the same regfile port names
Luke Kenneth Casson Leighton [Thu, 11 Nov 2021 10:29:19 +0000 (10:29 +0000)]
invert numbering on CR HDLState.get_crregs
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 19:42:53 +0000 (19:42 +0000)]
update store data reg 10 to 0xfe in virtmode mmu test
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 19:09:14 +0000 (19:09 +0000)]
remove read of MSR, it is done by passing through PowerDecoder2
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 19:01:40 +0000 (19:01 +0000)]
allow MSR to be set in StateRegs in test_core.py
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 18:38:08 +0000 (18:38 +0000)]
add $Display of oper_r.msr in LDSTCompUnit
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 18:26:15 +0000 (18:26 +0000)]
whitespace
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 18:22:09 +0000 (18:22 +0000)]
morph regfiles to add hazard vector make_vecs function
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 18:20:07 +0000 (18:20 +0000)]
add fetch of MSR in LD/ST pipe_data
Tobias Platen [Wed, 10 Nov 2021 18:20:06 +0000 (19:20 +0100)]
add debug output for msr_pr
Tobias Platen [Wed, 10 Nov 2021 17:58:18 +0000 (18:58 +0100)]
Merge branch 'master' of ssh://git.libre-riscv.org:922/soc
Tobias Platen [Wed, 10 Nov 2021 17:57:57 +0000 (18:57 +0100)]
test testcase for exception
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 13:41:49 +0000 (13:41 +0000)]
make core busy_o part of the CoreOutput data structure
the FSM TestIssuer can use this to detect not to send anything to it
and the InOrderIssuer can safely ignore it as long as it takes care
of RaW hazards
Luke Kenneth Casson Leighton [Wed, 10 Nov 2021 13:23:44 +0000 (13:23 +0000)]
add a "fu_found" signal to core, which allows for an indicator that
no Function Unit (no Reservation Station) is currently available for this
instruction
Tobias Platen [Tue, 9 Nov 2021 19:43:00 +0000 (20:43 +0100)]
Merge branch 'master' of ssh://git.libre-riscv.org:922/soc
Tobias Platen [Tue, 9 Nov 2021 19:42:29 +0000 (20:42 +0100)]
test_issuer_mmu.py: add case_5_allsprs
Luke Kenneth Casson Leighton [Tue, 9 Nov 2021 17:49:30 +0000 (17:49 +0000)]
add core instruction-issue PriorityPickers
this got more complicated than expected. it should have been dead easy:
* identify each type of ALU (Logical, Shift, Mul)
* create a list per type of ALU of all Function Units providing that Function
* create a PriorityPicker for each type of ALU
* pick one
* shove that into the "issue" of the picked Function Unit.
where that went wrong was nothing to do with issue, it was to do with the
selection of the register files.
a return result from connect_instruction is a dictionary of "enabled"
bits, one per FU, giving permission for that FU to perform read/write
access to regfile ports.
however with the fu_bitdict changing from "continuously-enabled" over
to "single-pulse indicating issue", the read/write ports were no longer
requested.
to fix this, an *additional* dictionary was created: fu_selected, which
has the OR of the FU issue *and* the FU "busy" signal, which will be
HI for the whole time that the FU is active (even including the first
issue cycle: normally busy_o only goes active one cycle *after* issue
is blipped)
thus in theory, although in practice it will not happen, the read/write
ports could be requested immediately the instruction is issued.
Luke Kenneth Casson Leighton [Tue, 9 Nov 2021 12:50:32 +0000 (12:50 +0000)]
comments
Luke Kenneth Casson Leighton [Tue, 9 Nov 2021 12:49:54 +0000 (12:49 +0000)]
core.py: create a dictionary of lists of Function Units capable of
dealing with a particular instruction (by power_enums Function: ALU, MMU, DIV
LOGICAL etc.)
Luke Kenneth Casson Leighton [Tue, 9 Nov 2021 12:38:25 +0000 (12:38 +0000)]
create function core conect_satellite_decoders
just moving code from elaborate to make it clearer
Luke Kenneth Casson Leighton [Tue, 9 Nov 2021 00:39:04 +0000 (00:39 +0000)]
add cancel in to alu_ok / alu_valid in LDSTCompUnit
Luke Kenneth Casson Leighton [Tue, 9 Nov 2021 00:34:42 +0000 (00:34 +0000)]
rename LDSTCompUnit cancel to canceln (because it is active low)
Luke Kenneth Casson Leighton [Tue, 9 Nov 2021 00:27:05 +0000 (00:27 +0000)]
whoops must remember to do rdmaskn on LDSTCompUnit as well
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 23:54:02 +0000 (23:54 +0000)]
remove unit test that is unfinished
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 23:48:05 +0000 (23:48 +0000)]
shorter way of getting FU busy signals
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 23:42:21 +0000 (23:42 +0000)]
MultiCompUnit fixed to not need rdmask to be sustained indefinitely
only needs to be set on input (when issue is raised)
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 23:38:58 +0000 (23:38 +0000)]
in MultiCompUnit, put rdmaskn into src latch rather than OR in src release
this allows rdmaskn to be set just once rather than left permanently
on. but, it does mean that the internal src register latches end up
having the wrong (previous) output, which then bleed through into
ALUs.
to stop that, the src data latches are all set to zero if the CompUnit
is not busy
Tobias Platen [Mon, 8 Nov 2021 20:02:07 +0000 (21:02 +0100)]
mmu unit test working again
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 16:37:51 +0000 (16:37 +0000)]
remove unused variable
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 16:34:57 +0000 (16:34 +0000)]
code comments
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 16:17:26 +0000 (16:17 +0000)]
comments
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 14:36:09 +0000 (14:36 +0000)]
remove issue_i from core, use i_valid instead to decide when to issue
converting core to Pipeline API
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 14:19:14 +0000 (14:19 +0000)]
move "exception happened" detection from TestIssuer to Core
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 14:14:43 +0000 (14:14 +0000)]
use p.i_valid in core instead of explicit signal ivalid_i
converting core to Pipeline API
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 14:08:05 +0000 (14:08 +0000)]
use Pipeline API o_ready instead of explicit core busy_o signal
converting core.py to Pipeline API
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 13:59:16 +0000 (13:59 +0000)]
convert core.py to Pipeline API, deriving from ControlBase
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 13:45:10 +0000 (13:45 +0000)]
remove unneeded imports
Luke Kenneth Casson Leighton [Mon, 8 Nov 2021 11:12:40 +0000 (11:12 +0000)]
move simple core input and output data to in/out data structures
Luke Kenneth Casson Leighton [Sun, 7 Nov 2021 22:39:24 +0000 (22:39 +0000)]
make FSMDivCoreStage properly conform to Stage API
added ispec and ospec functions, no need to explicitly set
self.p_i_data and self.n.o_data because that is the job of the Stage API
Luke Kenneth Casson Leighton [Sun, 7 Nov 2021 15:40:17 +0000 (15:40 +0000)]
add hazard vectors to Regfiles
the reason for adding it to Regfiles is because both In-Order and OoO
need global hazard vectors.
in the case of In-Order the hazard vector bits are set directly by the
Issue Engine.
in the case of Out-of-Order the vector bits are set by way of an
amalgamation (Great Big Or Gate) of the columns from the DMs
in either case the vectors are needed, so might as well be added to Regfiles
Luke Kenneth Casson Leighton [Sun, 7 Nov 2021 15:17:32 +0000 (15:17 +0000)]
add quick test of regfiles to output rtlil
Luke Kenneth Casson Leighton [Sun, 7 Nov 2021 13:45:41 +0000 (13:45 +0000)]
switch over to single-entry (num_rows=1) ReservationStation2 based
MultiCompUnit
this is effectively identical to FunctionUnitBaseSingle behaviour except
that now there is the possibility of increasing the number of rows
(number of "fronts" to the ALU) to 2 or greater
DIV FSM, Trap and SPR as well as LDST have been left alone
Luke Kenneth Casson Leighton [Sun, 7 Nov 2021 13:40:33 +0000 (13:40 +0000)]
for some reason mul test cases had not been added to test_issuer
Luke Kenneth Casson Leighton [Sun, 7 Nov 2021 12:57:12 +0000 (12:57 +0000)]
adding an FSM-based MultiCompUnit test (does not work yet)
Luke Kenneth Casson Leighton [Sun, 7 Nov 2021 12:29:53 +0000 (12:29 +0000)]
remove some of the uses of wrmask (redundant)
https://bugs.libre-soc.org/show_bug.cgi?id=742
in MultiCompUnit, wrmask is the amalgamation of the incoming "data ok"
signals from an ALU. ALUs are given the *option* to write to registers:
they are not told, "you absolutely have to write to this register"
but of course for Hazard purposes, the MultiCompUnit has to be told
to "hold" off the write until such time as the ALU can determine whether
it has anything to write
it _should_ now be possible for the ReservationStation version of
FunctionUnit to use combinatorial-setting of write data+ok
Tobias Platen [Sat, 6 Nov 2021 14:14:47 +0000 (15:14 +0100)]
update test_issuer_mmu.py testcase, add needed debug outputs
Tobias Platen [Fri, 5 Nov 2021 20:18:53 +0000 (21:18 +0100)]
tlbie, mtspr and mfspr test cases
Tobias Platen [Fri, 5 Nov 2021 19:15:16 +0000 (20:15 +0100)]
add mmu/dcache unit test
Luke Kenneth Casson Leighton [Thu, 4 Nov 2021 19:11:28 +0000 (19:11 +0000)]
use ReservationStations2 (disabled for now)
Luke Kenneth Casson Leighton [Thu, 4 Nov 2021 19:07:37 +0000 (19:07 +0000)]
write-ok is expected to stay valid *after* being set,
working out what is going on here
Luke Kenneth Casson Leighton [Thu, 4 Nov 2021 18:47:36 +0000 (18:47 +0000)]
add name to write pick on core
Luke Kenneth Casson Leighton [Thu, 4 Nov 2021 18:05:33 +0000 (18:05 +0000)]
fix missing naming ready_i -> i_ready
Tobias Platen [Wed, 3 Nov 2021 17:37:35 +0000 (18:37 +0100)]
cleanup fsm
Tobias Platen [Wed, 3 Nov 2021 17:34:57 +0000 (18:34 +0100)]
loadstore.py: add Display statement on SPR change
Tobias Platen [Wed, 3 Nov 2021 16:54:12 +0000 (17:54 +0100)]
add first tlbie test case