def __init__(self, pspec):
+ # test if microwatt compatibility is to be enabled
+ self.microwatt_compat = (hasattr(pspec, "microwatt_compat") and
+ (pspec.microwatt_compat == True))
+ self.alt_reset = Signal(reset_less=True) # not connected yet (microwatt)
+ # test if fabric compatibility is to be enabled
+ self.fabric_compat = (hasattr(pspec, "fabric_compat") and
+ (pspec.fabric_compat == True))
+
+ if self.microwatt_compat or self.fabric_compat:
+
+ if hasattr(pspec, "microwatt_old"):
+ self.microwatt_old = pspec.microwatt_old
+ else:
+ self.microwatt_old = True # PLEASE DO NOT ALTER THIS
+
+ if hasattr(pspec, "microwatt_debug"):
+ self.microwatt_debug = pspec.microwatt_debug
+ else:
+ self.microwatt_debug = True # set to False when using an FPGA
+
# test is SVP64 is to be enabled
self.svp64_en = hasattr(pspec, "svp64") and (pspec.svp64 == True)
self.allow_overlap = (hasattr(pspec, "allow_overlap") and
(pspec.allow_overlap == True))
+ # and get the core domain
+ self.core_domain = "coresync"
+ if (hasattr(pspec, "core_domain") and
+ isinstance(pspec.core_domain, str)):
+ self.core_domain = pspec.core_domain
+
# JTAG interface. add this right at the start because if it's
# added it *modifies* the pspec, by adding enable/disable signals
# for parts of the rest of the core
self.jtag_en = hasattr(pspec, "debug") and pspec.debug == 'jtag'
- self.dbg_domain = "sync" # sigh "dbgsunc" too problematic
- # self.dbg_domain = "dbgsync" # domain for DMI/JTAG clock
+ #self.dbg_domain = "sync" # sigh "dbgsunc" too problematic
+ self.dbg_domain = "dbgsync" # domain for DMI/JTAG clock
if self.jtag_en:
- # XXX MUST keep this up-to-date with litex, and
+ # XXX MUST keep this up-to-date with fabric, and
# soc-cocotb-sim, and err.. all needs sorting out, argh
subset = ['uart',
'mtwi',
self.xics_icp = XICS_ICP()
self.xics_ics = XICS_ICS()
self.int_level_i = self.xics_ics.int_level_i
+ else:
+ self.ext_irq = Signal()
# add GPIO peripheral?
self.gpio = hasattr(pspec, "gpio") and pspec.gpio == True
# main instruction core. suitable for prototyping / demo only
self.core = core = NonProductionCore(pspec)
- self.core_rst = ResetSignal("coresync")
+ self.core_rst = ResetSignal(self.core_domain)
# instruction decoder. goes into Trap Record
#pdecode = create_pdecode()
# DMI interface
self.dbg = CoreDebug()
+ self.dbg_rst_i = Signal(reset_less=True)
# instruction go/monitor
self.pc_o = Signal(64, reset_less=True)
self.state_r_pc = staterf.r_ports['cia'] # PC rd
self.state_r_sv = staterf.r_ports['sv'] # SVSTATE rd
- self.state_w_msr = staterf.w_ports['msr'] # MSR wr
+ self.state_w_msr = staterf.w_ports['d_wr2'] # MSR wr
self.state_w_pc = staterf.w_ports['d_wr1'] # PC wr
self.state_w_sv = staterf.w_ports['sv'] # SVSTATE wr
# DMI interface access
intrf = self.core.regs.rf['int']
+ fastrf = self.core.regs.rf['fast']
crrf = self.core.regs.rf['cr']
xerrf = self.core.regs.rf['xer']
- self.int_r = intrf.r_ports['dmi'] # INT read
- self.cr_r = crrf.r_ports['full_cr_dbg'] # CR read
- self.xer_r = xerrf.r_ports['full_xer'] # XER read
+ self.int_r = intrf.r_ports['dmi'] # INT DMI read
+ self.cr_r = crrf.r_ports['full_cr_dbg'] # CR DMI read
+ self.xer_r = xerrf.r_ports['full_xer'] # XER DMI read
+ self.fast_r = fastrf.r_ports['dmi'] # FAST DMI read
if self.svp64_en:
# for predication
# hack method of keeping an eye on whether branch/trap set the PC
self.state_nia = self.core.regs.rf['state'].w_ports['nia']
self.state_nia.wen.name = 'state_nia_wen'
+ # and whether SPR pipeline sets DEC or TB (fu/spr/main_stage.py)
+ self.state_spr = self.core.regs.rf['state'].w_ports['state1']
# pulse to synchronize the simulator at instruction end
self.insn_done = Signal()
self.srcmask = Signal(64)
self.dstmask = Signal(64)
+ # sigh, the wishbone addresses are not wishbone-compliant
+ # in old versions of microwatt, tplaten_3d_game is a new one
+ if self.microwatt_compat or self.fabric_compat:
+ self.ibus_adr = Signal(32, name='wishbone_insn_out.adr')
+ self.dbus_adr = Signal(32, name='wishbone_data_out.adr')
+
+ # add an output of the PC and instruction, and whether it was requested
+ # this is for verilator debug purposes
+ if self.microwatt_compat or self.fabric_compat:
+ self.nia = Signal(64)
+ self.msr_o = Signal(64)
+ self.nia_req = Signal(1)
+ self.insn = Signal(32)
+ self.ldst_req = Signal(1)
+ self.ldst_addr = Signal(1)
+
+ # for pausing dec/tb during an SPR pipeline event, this
+ # ensures that an SPR write (mtspr) to TB or DEC does not
+ # get overwritten by the DEC/TB FSM
+ self.pause_dec_tb = Signal()
+
def setup_peripherals(self, m):
comb, sync = m.d.comb, m.d.sync
# but NOT its reset signal. to cope with this, set every single
# submodule explicitly in coresync domain, debug and JTAG
# in their own one but using *external* reset.
- csd = DomainRenamer("coresync")
+ csd = DomainRenamer(self.core_domain)
dbd = DomainRenamer(self.dbg_domain)
- m.submodules.core = core = csd(self.core)
+ if self.microwatt_compat or self.fabric_compat:
+ m.submodules.core = core = self.core
+ else:
+ m.submodules.core = core = csd(self.core)
+
# this _so_ needs sorting out. ICache is added down inside
# LoadStore1 and is already a submodule of LoadStore1
if not isinstance(self.imem, ICache):
m.submodules.imem = imem = csd(self.imem)
+
+ # set up JTAG Debug Module (in correct domain)
m.submodules.dbg = dbg = dbd(self.dbg)
if self.jtag_en:
m.submodules.jtag = jtag = dbd(self.jtag)
# see https://bugs.libre-soc.org/show_bug.cgi?id=499
sync += dbg.dmi.connect_to(jtag.dmi)
+ # fixup the clocks in microwatt-compat mode (but leave resets alone
+ # so that microwatt soc.vhdl can pull a reset on the core or DMI
+ # can do it, just like in TestIssuer)
+ if self.microwatt_compat or self.fabric_compat:
+ intclk = ClockSignal(self.core_domain)
+ dbgclk = ClockSignal(self.dbg_domain)
+ if self.core_domain != 'sync':
+ comb += intclk.eq(ClockSignal())
+ if self.dbg_domain != 'sync':
+ comb += dbgclk.eq(ClockSignal())
+
+ # if using old version of microwatt
+ # drop the first 3 bits of the incoming wishbone addresses
+ if self.microwatt_compat or self.fabric_compat:
+ ibus = self.imem.ibus
+ dbus = self.core.l0.cmpi.wb_bus()
+ if self.microwatt_old:
+ comb += self.ibus_adr.eq(Cat(Const(0, 3), ibus.adr))
+ comb += self.dbus_adr.eq(Cat(Const(0, 3), dbus.adr))
+ else:
+ comb += self.ibus_adr.eq(ibus.adr)
+ comb += self.dbus_adr.eq(dbus.adr)
+ if self.microwatt_debug:
+ # microwatt verilator debug purposes
+ pi = self.core.l0.cmpi.pi.pi
+ comb += self.ldst_req.eq(pi.addr_ok_o)
+ comb += self.ldst_addr.eq(pi.addr)
+
cur_state = self.cur_state
- # 4x 4k SRAM blocks. these simply "exist", they get routed in litex
+ # 4x 4k SRAM blocks. these simply "exist", they get routed in fabric
if self.sram4x4k:
for i, sram in enumerate(self.sram4k):
m.submodules["sram4k_%d" % i] = csd(sram)
m.submodules.xics_ics = ics = csd(self.xics_ics)
comb += icp.ics_i.eq(ics.icp_o) # connect ICS to ICP
sync += cur_state.eint.eq(icp.core_irq_o) # connect ICP to core
+ else:
+ sync += cur_state.eint.eq(self.ext_irq) # connect externally
# GPIO test peripheral
if self.gpio:
m.submodules.simple_gpio = simple_gpio = csd(self.simple_gpio)
# connect one GPIO output to ICS bit 15 (like in microwatt soc.vhdl)
- # XXX causes litex ECP5 test to get wrong idea about input and output
+ # XXX causes fabric ECP5 test to get wrong idea about input and output
# (but works with verilator sim *sigh*)
# if self.gpio and self.xics:
# comb += self.int_level_i[15].eq(simple_gpio.gpio_o[0])
if self.svp64_en:
m.submodules.svp64 = svp64 = csd(self.svp64)
- # convenience
- dmi, d_reg, d_cr, d_xer, = dbg.dmi, dbg.d_gpr, dbg.d_cr, dbg.d_xer
- intrf = self.core.regs.rf['int']
-
# clock delay power-on reset
cd_por = ClockDomain(reset_less=True)
cd_sync = ClockDomain()
- core_sync = ClockDomain("coresync")
- m.domains += cd_por, cd_sync, core_sync
+ m.domains += cd_por, cd_sync
+ core_sync = ClockDomain(self.core_domain)
+ if self.core_domain != "sync":
+ m.domains += core_sync
if self.dbg_domain != "sync":
dbg_sync = ClockDomain(self.dbg_domain)
m.domains += dbg_sync
+ # create a delay, but remember it is in the power-on-reset clock domain!
ti_rst = Signal(reset_less=True)
delay = Signal(range(4), reset=3)
+ stop_delay = Signal(range(16), reset=5)
with m.If(delay != 0):
- m.d.por += delay.eq(delay - 1)
+ m.d.por += delay.eq(delay - 1) # decrement... in POR domain!
+ with m.If(stop_delay != 0):
+ m.d.por += stop_delay.eq(stop_delay - 1) # likewise
comb += cd_por.clk.eq(ClockSignal())
# power-on reset delay
- core_rst = ResetSignal("coresync")
- comb += ti_rst.eq(delay != 0 | dbg.core_rst_o | ResetSignal())
- comb += core_rst.eq(ti_rst)
+ core_rst = ResetSignal(self.core_domain)
+ if self.core_domain != "sync":
+ comb += ti_rst.eq(delay != 0 | dbg.core_rst_o | ResetSignal())
+ comb += core_rst.eq(ti_rst)
+ else:
+ with m.If(delay != 0 | dbg.core_rst_o):
+ comb += core_rst.eq(1)
+ with m.If(stop_delay != 0):
+ # run DMI core-stop as well but on an extra couple of cycles
+ comb += dbg.core_stopped_i.eq(1)
- # debug clock is same as coresync, but reset is *main external*
+ # connect external reset signal to DMI Reset
if self.dbg_domain != "sync":
dbg_rst = ResetSignal(self.dbg_domain)
- comb += dbg_rst.eq(ResetSignal())
+ comb += dbg_rst.eq(self.dbg_rst_i)
# busy/halted signals from core
core_busy_o = ~core.p.o_ready | core.n.o_data.busy_o # core is busy
comb += pdecode2.dec.bigendian.eq(self.core_bigendian_i)
# temporary hack: says "go" immediately for both address gen and ST
+ # XXX: st.go_i is set to 1 cycle delay to reduce combinatorial chains
l0 = core.l0
ldst = core.fus.fus['ldst0']
st_go_edge = rising_edge(m, ldst.st.rel_o)
# link addr-go direct to rel
m.d.comb += ldst.ad.go_i.eq(ldst.ad.rel_o)
- m.d.comb += ldst.st.go_i.eq(st_go_edge) # link store-go to rising rel
+ m.d.sync += ldst.st.go_i.eq(st_go_edge) # link store-go to rising rel
def do_dmi(self, m, dbg):
"""deals with DMI debug requests
comb = m.d.comb
sync = m.d.sync
dmi, d_reg, d_cr, d_xer, = dbg.dmi, dbg.d_gpr, dbg.d_cr, dbg.d_xer
+ d_fast = dbg.d_fast
intrf = self.core.regs.rf['int']
+ fastrf = self.core.regs.rf['fast']
with m.If(d_reg.req): # request for regfile access being made
# TODO: error-check this
comb += d_reg.data.eq(self.int_r.o_data)
comb += d_reg.ack.eq(1)
+ # fast regfile
+ with m.If(d_fast.req): # request for regfile access being made
+ if fastrf.unary:
+ comb += self.fast_r.ren.eq(1 << d_fast.addr)
+ else:
+ comb += self.fast_r.addr.eq(d_fast.addr)
+ comb += self.fast_r.ren.eq(1)
+ d_fast_delay = Signal()
+ sync += d_fast_delay.eq(d_fast.req)
+ with m.If(d_fast_delay):
+ # data arrives one clock later
+ comb += d_fast.data.eq(self.fast_r.o_data)
+ comb += d_fast.ack.eq(1)
+
# sigh same thing for CR debug
with m.If(d_cr.req): # request for regfile access being made
comb += self.cr_r.ren.eq(0b11111111) # enable all
value to DEC, however the regfile has "passthrough" on it so this
*should* be ok.
- see v3.0B p1097-1099 for Timeer Resource and p1065 and p1076
+ see v3.0B p1097-1099 for Timer Resource and p1065 and p1076
"""
comb, sync = m.d.comb, m.d.sync
- fast_rf = self.core.regs.rf['fast']
- fast_r_dectb = fast_rf.r_ports['issue'] # DEC/TB
- fast_w_dectb = fast_rf.w_ports['issue'] # DEC/TB
+ state_rf = self.core.regs.rf['state']
+ state_r_dectb = state_rf.r_ports['issue'] # DEC/TB
+ state_w_dectb = state_rf.w_ports['issue'] # DEC/TB
+
with m.FSM() as fsm:
# initiates read of current DEC
with m.State("DEC_READ"):
- comb += fast_r_dectb.addr.eq(FastRegs.DEC)
- comb += fast_r_dectb.ren.eq(1)
- m.next = "DEC_WRITE"
+ comb += state_r_dectb.ren.eq(1<<StateRegs.DEC)
+ with m.If(~self.pause_dec_tb):
+ m.next = "DEC_WRITE"
# waits for DEC read to arrive (1 cycle), updates with new value
+ # respects if dec/tb writing has been paused
with m.State("DEC_WRITE"):
- new_dec = Signal(64)
- # TODO: MSR.LPCR 32-bit decrement mode
- comb += new_dec.eq(fast_r_dectb.o_data - 1)
- comb += fast_w_dectb.addr.eq(FastRegs.DEC)
- comb += fast_w_dectb.wen.eq(1)
- comb += fast_w_dectb.i_data.eq(new_dec)
- sync += spr_dec.eq(new_dec) # copy into cur_state for decoder
- m.next = "TB_READ"
+ with m.If(self.pause_dec_tb):
+ # if paused, return to reading
+ m.next = "DEC_READ"
+ with m.Else():
+ new_dec = Signal(64)
+ # TODO: MSR.LPCR 32-bit decrement mode
+ comb += new_dec.eq(state_r_dectb.o_data - 1)
+ comb += state_w_dectb.wen.eq(1<<StateRegs.DEC)
+ comb += state_w_dectb.i_data.eq(new_dec)
+ # copy to cur_state for decoder, for an interrupt
+ sync += spr_dec.eq(new_dec)
+ m.next = "TB_READ"
# initiates read of current TB
with m.State("TB_READ"):
- comb += fast_r_dectb.addr.eq(FastRegs.TB)
- comb += fast_r_dectb.ren.eq(1)
- m.next = "TB_WRITE"
+ comb += state_r_dectb.ren.eq(1<<StateRegs.TB)
+ with m.If(~self.pause_dec_tb):
+ m.next = "TB_WRITE"
# waits for read TB to arrive, initiates write of current TB
+ # respects if dec/tb writing has been paused
with m.State("TB_WRITE"):
- new_tb = Signal(64)
- comb += new_tb.eq(fast_r_dectb.o_data + 1)
- comb += fast_w_dectb.addr.eq(FastRegs.TB)
- comb += fast_w_dectb.wen.eq(1)
- comb += fast_w_dectb.i_data.eq(new_tb)
- m.next = "DEC_READ"
+ with m.If(self.pause_dec_tb):
+ # if paused, return to reading
+ m.next = "TB_READ"
+ with m.Else():
+ new_tb = Signal(64)
+ comb += new_tb.eq(state_r_dectb.o_data + 1)
+ comb += state_w_dectb.wen.eq(1<<StateRegs.TB)
+ comb += state_w_dectb.i_data.eq(new_tb)
+ m.next = "DEC_READ"
return m
# reset current state if core reset requested
with m.If(core_rst):
m.d.sync += self.cur_state.eq(0)
+ # and, sigh, set configured values, which are also done in regfile
+ # XXX ??? what the hell is the shift for??
+ m.d.sync += self.cur_state.pc.eq(self.core.pc_at_reset)
+ m.d.sync += self.cur_state.msr.eq(self.core.msr_at_reset)
+
+ # check halted condition: requested PC to execute matches DMI stop addr
+ # and immediately stop. address of 0xffff_ffff_ffff_ffff can never
+ # match
+ halted = Signal()
+ comb += halted.eq(dbg.stop_addr_o == dbg.state.pc)
+ with m.If(halted):
+ comb += dbg.core_stopped_i.eq(1)
+ comb += dbg.terminate_i.eq(1)
# PC and instruction from I-Memory
comb += self.pc_o.eq(cur_state.pc)
comb += self.state_w_sv.i_data.eq(self.new_svstate)
sync += self.sv_changed.eq(1)
+ # start renaming some of the ports to match microwatt
+ if self.microwatt_compat or self.fabric_compat:
+ self.core.o.core_terminate_o.name = "terminated_out"
+ # names of DMI interface
+ self.dbg.dmi.addr_i.name = 'dmi_addr'
+ self.dbg.dmi.din.name = 'dmi_din'
+ self.dbg.dmi.dout.name = 'dmi_dout'
+ self.dbg.dmi.req_i.name = 'dmi_req'
+ self.dbg.dmi.we_i.name = 'dmi_wr'
+ self.dbg.dmi.ack_o.name = 'dmi_ack'
+ # wishbone instruction bus
+ ibus = self.imem.ibus
+ if self.microwatt_compat:
+ ibus.adr.name = 'wishbone_insn_out.adr'
+ ibus.dat_w.name = 'wishbone_insn_out.dat'
+ ibus.sel.name = 'wishbone_insn_out.sel'
+ ibus.cyc.name = 'wishbone_insn_out.cyc'
+ ibus.stb.name = 'wishbone_insn_out.stb'
+ ibus.we.name = 'wishbone_insn_out.we'
+ ibus.dat_r.name = 'wishbone_insn_in.dat'
+ ibus.ack.name = 'wishbone_insn_in.ack'
+ ibus.stall.name = 'wishbone_insn_in.stall'
+ # wishbone data bus
+ dbus = self.core.l0.cmpi.wb_bus()
+ if self.microwatt_compat:
+ dbus.adr.name = 'wishbone_data_out.adr'
+ dbus.dat_w.name = 'wishbone_data_out.dat'
+ dbus.sel.name = 'wishbone_data_out.sel'
+ dbus.cyc.name = 'wishbone_data_out.cyc'
+ dbus.stb.name = 'wishbone_data_out.stb'
+ dbus.we.name = 'wishbone_data_out.we'
+ dbus.dat_r.name = 'wishbone_data_in.dat'
+ dbus.ack.name = 'wishbone_data_in.ack'
+ dbus.stall.name = 'wishbone_data_in.stall'
+
return m
def __iter__(self):
return list(self)
def external_ports(self):
+ if self.microwatt_compat or self.fabric_compat:
+ if self.fabric_compat:
+ ports = [self.core.o.core_terminate_o,
+ self.alt_reset, # not connected yet
+ self.nia, self.insn, self.nia_req, self.msr_o,
+ self.ldst_req, self.ldst_addr,
+ ClockSignal(),
+ ResetSignal(),
+ ]
+ else:
+ ports = [self.core.o.core_terminate_o,
+ self.ext_irq,
+ self.alt_reset, # not connected yet
+ self.nia, self.insn, self.nia_req, self.msr_o,
+ self.ldst_req, self.ldst_addr,
+ ClockSignal(),
+ ResetSignal(),
+ ]
+ ports += list(self.dbg.dmi.ports())
+ # for dbus/ibus microwatt, exclude err btw and cti
+ for name, sig in self.imem.ibus.fields.items():
+ if name not in ['err', 'bte', 'cti', 'adr']:
+ ports.append(sig)
+ for name, sig in self.core.l0.cmpi.wb_bus().fields.items():
+ if name not in ['err', 'bte', 'cti', 'adr']:
+ ports.append(sig)
+ # microwatt non-compliant with wishbone
+ ports.append(self.ibus_adr)
+ ports.append(self.dbus_adr)
+
+ if self.microwatt_compat:
+ # Ignore the remaining ports in microwatt compat mode
+ return ports
+
ports = self.pc_i.ports()
ports = self.msr_i.ports()
ports += [self.pc_o, self.memerr_o, self.core_bigendian_i, self.busy_o,
ports += list(self.xics_icp.bus.fields.values())
ports += list(self.xics_ics.bus.fields.values())
ports.append(self.int_level_i)
+ else:
+ ports.append(self.ext_irq)
if self.gpio:
ports += list(self.simple_gpio.bus.fields.values())
return list(self)
+class TestIssuerInternal(TestIssuerBase):
+ """TestIssuer - reads instructions from TestMemory and issues them
-# Fetch Finite State Machine.
-# WARNING: there are currently DriverConflicts but it's actually working.
-# TODO, here: everything that is global in nature, information from the
-# main TestIssuerInternal, needs to move to either ispec() or ospec().
-# not only that: TestIssuerInternal.imem can entirely move into here
-# because imem is only ever accessed inside the FetchFSM.
-class FetchFSM(ControlBase):
- def __init__(self, allow_overlap, svp64_en, imem, core_rst,
- pdecode2, cur_state,
- dbg, core, svstate, nia, is_svp64_mode):
- self.allow_overlap = allow_overlap
- self.svp64_en = svp64_en
- self.imem = imem
- self.core_rst = core_rst
- self.pdecode2 = pdecode2
- self.cur_state = cur_state
- self.dbg = dbg
- self.core = core
- self.svstate = svstate
- self.nia = nia
- self.is_svp64_mode = is_svp64_mode
-
- # set up pipeline ControlBase and allocate i/o specs
- # (unusual: normally done by the Pipeline API)
- super().__init__(stage=self)
- self.p.i_data, self.n.o_data = self.new_specs(None)
- self.i, self.o = self.p.i_data, self.n.o_data
-
- # next 3 functions are Stage API Compliance
- def setup(self, m, i):
- pass
-
- def ispec(self):
- return FetchInput()
-
- def ospec(self):
- return FetchOutput()
+ efficiency and speed is not the main goal here: functional correctness
+ and code clarity is. optimisations (which almost 100% interfere with
+ easy understanding) come later.
+ """
- def elaborate(self, platform):
+ def fetch_fsm(self, m, dbg, core, core_rst, nia, is_svp64_mode,
+ fetch_pc_o_ready, fetch_pc_i_valid,
+ fetch_insn_o_valid, fetch_insn_i_ready):
"""fetch FSM
this FSM performs fetch of raw instruction data, partial-decodes
it 32-bit at a time to detect SVP64 prefixes, and will optionally
read a 2nd 32-bit quantity if that occurs.
"""
- m = super().elaborate(platform)
-
- dbg = self.dbg
- core = self.core
- pc = self.i.pc
- msr = self.i.msr
- svstate = self.svstate
- nia = self.nia
- is_svp64_mode = self.is_svp64_mode
- fetch_pc_o_ready = self.p.o_ready
- fetch_pc_i_valid = self.p.i_valid
- fetch_insn_o_valid = self.n.o_valid
- fetch_insn_i_ready = self.n.i_ready
-
comb = m.d.comb
sync = m.d.sync
pdecode2 = self.pdecode2
cur_state = self.cur_state
- dec_opcode_o = pdecode2.dec.raw_opcode_in # raw opcode
+ dec_opcode_i = pdecode2.dec.raw_opcode_in # raw opcode
+ pc, msr, svstate = cur_state.pc, cur_state.msr, cur_state.svstate
# also note instruction fetch failed
if hasattr(core, "icache"):
# set priv / virt mode on I-Cache, sigh
if isinstance(self.imem, ICache):
comb += self.imem.i_in.priv_mode.eq(~msr[MSR.PR])
- comb += self.imem.i_in.virt_mode.eq(msr[MSR.DR])
+ comb += self.imem.i_in.virt_mode.eq(msr[MSR.IR]) # Instr. Redir (VM)
with m.FSM(name='fetch_fsm'):
+ # allow fetch to not run at startup due to I-Cache reset not
+ # having time to settle. power-on-reset holds dbg.core_stopped_i
+ with m.State("PRE_IDLE"):
+ with m.If(~dbg.core_stopped_i & ~dbg.core_stop_o & ~core_rst):
+ m.next = "IDLE"
+
# waiting (zzz)
with m.State("IDLE"):
- with m.If(~dbg.stopping_o & ~fetch_failed):
+ # fetch allowed if not failed and stopped but not stepping
+ # (see dmi.py for how core_stop_o is generated)
+ with m.If(~fetch_failed & ~dbg.core_stop_o):
comb += fetch_pc_o_ready.eq(1)
- with m.If(fetch_pc_i_valid & ~fetch_failed):
+ with m.If(fetch_pc_i_valid & ~pdecode2.instr_fault
+ & ~dbg.core_stop_o):
# instruction allowed to go: start by reading the PC
# capture the PC and also drop it into Insn Memory
# we have joined a pair of combinatorial memory
comb += self.imem.a_pc_i.eq(pc)
comb += self.imem.a_i_valid.eq(1)
comb += self.imem.f_i_valid.eq(1)
- # transfer state to output
- sync += cur_state.pc.eq(pc)
- sync += cur_state.svstate.eq(svstate) # and svstate
- sync += cur_state.msr.eq(msr) # and msr
-
m.next = "INSN_READ" # move to "wait for bus" phase
# dummy pause to find out why simulation is not keeping up
with m.State("INSN_READ"):
- if self.allow_overlap:
- stopping = dbg.stopping_o
- else:
- stopping = Const(0)
+ # when using "single-step" mode, checking dbg.stopping_o
+ # prevents progress. allow fetch to proceed once started
+ stopping = Const(0)
+ #if self.allow_overlap:
+ # stopping = dbg.stopping_o
with m.If(stopping):
# stopping: jump back to idle
m.next = "IDLE"
with m.Else():
- with m.If(self.imem.f_busy_o & ~fetch_failed): # zzz...
+ with m.If(self.imem.f_busy_o &
+ ~pdecode2.instr_fault): # zzz...
# busy but not fetch failed: stay in wait-read
+ comb += self.imem.a_pc_i.eq(pc)
comb += self.imem.a_i_valid.eq(1)
comb += self.imem.f_i_valid.eq(1)
with m.Else():
# not busy (or fetch failed!): instruction fetched
# when fetch failed, the instruction gets ignored
# by the decoder
- insn = get_insn(self.imem.f_instr_o, cur_state.pc)
+ if hasattr(core, "icache"):
+ # blech, icache returns actual instruction
+ insn = self.imem.f_instr_o
+ else:
+ # but these return raw memory
+ insn = get_insn(self.imem.f_instr_o, cur_state.pc)
if self.svp64_en:
svp64 = self.svp64
# decode the SVP64 prefix, if any
with m.If(~svp64.is_svp64_mode):
# with no prefix, store the instruction
# and hand it directly to the next FSM
- sync += dec_opcode_o.eq(insn)
+ sync += dec_opcode_i.eq(insn)
m.next = "INSN_READY"
with m.Else():
# fetch the rest of the instruction from memory
else:
# not SVP64 - 32-bit only
sync += nia.eq(cur_state.pc + 4)
- sync += dec_opcode_o.eq(insn)
+ sync += dec_opcode_i.eq(insn)
+ if self.microwatt_compat or self.fabric_compat:
+ # for verilator debug purposes
+ comb += self.insn.eq(insn)
+ comb += self.nia.eq(cur_state.pc)
+ comb += self.msr_o.eq(cur_state.msr)
+ comb += self.nia_req.eq(1)
m.next = "INSN_READY"
with m.State("INSN_READ2"):
comb += self.imem.f_i_valid.eq(1)
with m.Else():
# not busy: instruction fetched
- insn = get_insn(self.imem.f_instr_o, cur_state.pc+4)
- sync += dec_opcode_o.eq(insn)
+ if hasattr(core, "icache"):
+ # blech, icache returns actual instruction
+ insn = self.imem.f_instr_o
+ else:
+ insn = get_insn(self.imem.f_instr_o, cur_state.pc+4)
+ sync += dec_opcode_i.eq(insn)
m.next = "INSN_READY"
# TODO: probably can start looking at pdecode2.rm_dec
# here or maybe even in INSN_READ state, if svp64_mode
with m.If(fetch_insn_i_ready):
m.next = "IDLE"
- # whatever was done above, over-ride it if core reset is held
- with m.If(self.core_rst):
- sync += nia.eq(0)
-
- return m
-
-
-class TestIssuerInternal(TestIssuerBase):
- """TestIssuer - reads instructions from TestMemory and issues them
-
- efficiency and speed is not the main goal here: functional correctness
- and code clarity is. optimisations (which almost 100% interfere with
- easy understanding) come later.
- """
def fetch_predicate_fsm(self, m,
pred_insn_i_valid, pred_insn_o_ready,
fetch_failed = Const(0, 1)
flush_needed = False
+ sync += fetch_pc_i_valid.eq(0)
+
with m.FSM(name="issue_fsm"):
+ with m.State("PRE_IDLE"):
+ with m.If(~dbg.core_stop_o & ~core_rst):
+ m.next = "ISSUE_START"
+
# sync with the "fetch" phase which is reading the instruction
# at this point, there is no instruction running, that
# could inadvertently update the PC.
# wait on "core stop" release, before next fetch
# need to do this here, in case we are in a VL==0 loop
with m.If(~dbg.core_stop_o & ~core_rst):
- comb += fetch_pc_i_valid.eq(1) # tell fetch to start
+ sync += fetch_pc_i_valid.eq(1) # tell fetch to start
+ sync += cur_state.pc.eq(dbg.state.pc)
+ sync += cur_state.svstate.eq(dbg.state.svstate)
+ sync += cur_state.msr.eq(dbg.state.msr)
with m.If(fetch_pc_o_ready): # fetch acknowledged us
m.next = "INSN_WAIT"
with m.Else():
# wait for an instruction to arrive from Fetch
with m.State("INSN_WAIT"):
- if self.allow_overlap:
- stopping = dbg.stopping_o
- else:
- stopping = Const(0)
+ # when using "single-step" mode, checking dbg.stopping_o
+ # prevents progress. allow issue to proceed once started
+ stopping = Const(0)
+ #if self.allow_overlap:
+ # stopping = dbg.stopping_o
with m.If(stopping):
# stopping: jump back to idle
m.next = "ISSUE_START"
# handshake with execution FSM, move to "wait" once acknowledged
with m.State("INSN_EXECUTE"):
- comb += exec_insn_i_valid.eq(1) # trigger execute
- with m.If(exec_insn_o_ready): # execute acknowledged us
- m.next = "EXECUTE_WAIT"
+ # when using "single-step" mode, checking dbg.stopping_o
+ # prevents progress. allow execute to proceed once started
+ stopping = Const(0)
+ #if self.allow_overlap:
+ # stopping = dbg.stopping_o
+ with m.If(stopping):
+ # stopping: jump back to idle
+ m.next = "ISSUE_START"
+ if flush_needed:
+ # request the icache to stop asserting "failed"
+ comb += core.icache.flush_in.eq(1)
+ # stop instruction fault
+ sync += pdecode2.instr_fault.eq(0)
+ with m.Else():
+ comb += exec_insn_i_valid.eq(1) # trigger execute
+ with m.If(exec_insn_o_ready): # execute acknowledged us
+ m.next = "EXECUTE_WAIT"
with m.State("EXECUTE_WAIT"):
- # wait on "core stop" release, at instruction end
- # need to do this here, in case we are in a VL>1 loop
- with m.If(~dbg.core_stop_o & ~core_rst):
- comb += exec_pc_i_ready.eq(1)
- # see https://bugs.libre-soc.org/show_bug.cgi?id=636
- # the exception info needs to be blatted into
- # pdecode.ldst_exc, and the instruction "re-run".
- # when ldst_exc.happened is set, the PowerDecoder2
- # reacts very differently: it re-writes the instruction
- # with a "trap" (calls PowerDecoder2.trap()) which
- # will *overwrite* whatever was requested and jump the
- # PC to the exception address, as well as alter MSR.
- # nothing else needs to be done other than to note
- # the change of PC and MSR (and, later, SVSTATE)
- with m.If(exc_happened):
- mmu = core.fus.get_exc("mmu0")
- ldst = core.fus.get_exc("ldst0")
- if mmu is not None:
- with m.If(fetch_failed):
- # instruction fetch: exception is from MMU
- # reset instr_fault (highest priority)
- sync += pdecode2.ldst_exc.eq(mmu)
- sync += pdecode2.instr_fault.eq(0)
- if flush_needed:
- # request icache to stop asserting "failed"
- comb += core.icache.flush_in.eq(1)
- with m.If(~fetch_failed):
- # otherwise assume it was a LDST exception
- sync += pdecode2.ldst_exc.eq(ldst)
-
- with m.If(exec_pc_o_valid):
-
- # was this the last loop iteration?
- is_last = Signal()
- cur_vl = cur_state.svstate.vl
- comb += is_last.eq(next_srcstep == cur_vl)
-
- # return directly to Decode if Execute generated an
- # exception.
- with m.If(pdecode2.ldst_exc.happened):
- m.next = "DECODE_SV"
-
- # if MSR, PC or SVSTATE were changed by the previous
- # instruction, go directly back to Fetch, without
- # updating either MSR PC or SVSTATE
- with m.Elif(self.msr_changed | self.pc_changed |
- self.sv_changed):
- m.next = "ISSUE_START"
-
- # also return to Fetch, when no output was a vector
- # (regardless of SRCSTEP and VL), or when the last
- # instruction was really the last one of the VL loop
- with m.Elif((~pdecode2.loop_continue) | is_last):
- # before going back to fetch, update the PC state
- # register with the NIA.
- # ok here we are not reading the branch unit.
- # TODO: this just blithely overwrites whatever
- # pipeline updated the PC
- comb += self.state_w_pc.wen.eq(1 << StateRegs.PC)
- comb += self.state_w_pc.i_data.eq(nia)
- # reset SRCSTEP before returning to Fetch
- if self.svp64_en:
- with m.If(pdecode2.loop_continue):
- comb += new_svstate.srcstep.eq(0)
- comb += new_svstate.dststep.eq(0)
- comb += self.update_svstate.eq(1)
- else:
+ comb += exec_pc_i_ready.eq(1)
+ # see https://bugs.libre-soc.org/show_bug.cgi?id=636
+ # the exception info needs to be blatted into
+ # pdecode.ldst_exc, and the instruction "re-run".
+ # when ldst_exc.happened is set, the PowerDecoder2
+ # reacts very differently: it re-writes the instruction
+ # with a "trap" (calls PowerDecoder2.trap()) which
+ # will *overwrite* whatever was requested and jump the
+ # PC to the exception address, as well as alter MSR.
+ # nothing else needs to be done other than to note
+ # the change of PC and MSR (and, later, SVSTATE)
+ with m.If(exc_happened):
+ mmu = core.fus.get_exc("mmu0")
+ ldst = core.fus.get_exc("ldst0")
+ if mmu is not None:
+ with m.If(fetch_failed):
+ # instruction fetch: exception is from MMU
+ # reset instr_fault (highest priority)
+ sync += pdecode2.ldst_exc.eq(mmu)
+ sync += pdecode2.instr_fault.eq(0)
+ if flush_needed:
+ # request icache to stop asserting "failed"
+ comb += core.icache.flush_in.eq(1)
+ with m.If(~fetch_failed):
+ # otherwise assume it was a LDST exception
+ sync += pdecode2.ldst_exc.eq(ldst)
+
+ with m.If(exec_pc_o_valid):
+
+ # was this the last loop iteration?
+ is_last = Signal()
+ cur_vl = cur_state.svstate.vl
+ comb += is_last.eq(next_srcstep == cur_vl)
+
+ with m.If(pdecode2.instr_fault):
+ # reset instruction fault, try again
+ sync += pdecode2.instr_fault.eq(0)
+ m.next = "ISSUE_START"
+
+ # return directly to Decode if Execute generated an
+ # exception.
+ with m.Elif(pdecode2.ldst_exc.happened):
+ m.next = "DECODE_SV"
+
+ # if MSR, PC or SVSTATE were changed by the previous
+ # instruction, go directly back to Fetch, without
+ # updating either MSR PC or SVSTATE
+ with m.Elif(self.msr_changed | self.pc_changed |
+ self.sv_changed):
+ m.next = "ISSUE_START"
+
+ # also return to Fetch, when no output was a vector
+ # (regardless of SRCSTEP and VL), or when the last
+ # instruction was really the last one of the VL loop
+ with m.Elif((~pdecode2.loop_continue) | is_last):
+ # before going back to fetch, update the PC state
+ # register with the NIA.
+ # ok here we are not reading the branch unit.
+ # TODO: this just blithely overwrites whatever
+ # pipeline updated the PC
+ comb += self.state_w_pc.wen.eq(1 << StateRegs.PC)
+ comb += self.state_w_pc.i_data.eq(nia)
+ # reset SRCSTEP before returning to Fetch
+ if self.svp64_en:
+ with m.If(pdecode2.loop_continue):
comb += new_svstate.srcstep.eq(0)
comb += new_svstate.dststep.eq(0)
comb += self.update_svstate.eq(1)
- m.next = "ISSUE_START"
-
- # returning to Execute? then, first update SRCSTEP
- with m.Else():
- comb += new_svstate.srcstep.eq(next_srcstep)
- comb += new_svstate.dststep.eq(next_dststep)
+ else:
+ comb += new_svstate.srcstep.eq(0)
+ comb += new_svstate.dststep.eq(0)
comb += self.update_svstate.eq(1)
- # return to mask skip loop
- m.next = "PRED_SKIP"
+ m.next = "ISSUE_START"
+
+ # returning to Execute? then, first update SRCSTEP
+ with m.Else():
+ comb += new_svstate.srcstep.eq(next_srcstep)
+ comb += new_svstate.dststep.eq(next_dststep)
+ comb += self.update_svstate.eq(1)
+ # return to mask skip loop
+ m.next = "PRED_SKIP"
- with m.Else():
- comb += dbg.core_stopped_i.eq(1)
- if flush_needed:
- # request the icache to stop asserting "failed"
- comb += core.icache.flush_in.eq(1)
- # stop instruction fault
- sync += pdecode2.instr_fault.eq(0)
# check if svstate needs updating: if so, write it to State Regfile
with m.If(self.update_svstate):
comb = m.d.comb
sync = m.d.sync
+ dbg = self.dbg
pdecode2 = self.pdecode2
+ cur_state = self.cur_state
# temporaries
core_busy_o = core.n.o_data.busy_o # core is busy
# instruction started: must wait till it finishes
with m.State("INSN_ACTIVE"):
# note changes to MSR, PC and SVSTATE
- # XXX oops, really must monitor *all* State Regfile write
- # ports looking for changes!
with m.If(self.state_nia.wen & (1 << StateRegs.SVSTATE)):
sync += self.sv_changed.eq(1)
with m.If(self.state_nia.wen & (1 << StateRegs.MSR)):
sync += self.msr_changed.eq(1)
with m.If(self.state_nia.wen & (1 << StateRegs.PC)):
sync += self.pc_changed.eq(1)
+ # and note changes to DEC/TB, to be passed to DEC/TB FSM
+ with m.If(self.state_spr.wen & (1 << StateRegs.TB)):
+ comb += self.pause_dec_tb.eq(1)
+ # but also zero-out the cur_state DEC so that, on
+ # the next instruction, if it is "enable interrupt"
+ # the delay between the DEC/TB FSM reading and updating
+ # cur_state.dec doesn't trigger a spurious interrupt.
+ # the DEC/TB FSM will read the regfile and update to
+ # the correct value, so having cur_state.dec set to zero
+ # for a while is no big deal.
+ with m.If(self.state_spr.wen & (1 << StateRegs.DEC)):
+ comb += self.pause_dec_tb.eq(1)
+ sync += cur_state.dec.eq(0) # only needs top bit clear
with m.If(~core_busy_o): # instruction done!
comb += exec_pc_o_valid.eq(1)
with m.If(exec_pc_i_ready):
# there were *TWO* instructions:
# 1) the failed LDST 2) a TRAP.
with m.If(~pdecode2.ldst_exc.happened &
- ~fetch_failed):
+ ~pdecode2.instr_fault):
comb += self.insn_done.eq(1)
m.next = "INSN_START" # back to fetch
+ # terminate returns directly to INSN_START
+ with m.If(dbg.terminate_i):
+ # comb += self.insn_done.eq(1) - no because it's not
+ m.next = "INSN_START" # back to fetch
def elaborate(self, platform):
m = super().elaborate(platform)
nia = Signal(64)
# connect up debug signals
- comb += dbg.terminate_i.eq(core.o.core_terminate_o)
+ with m.If(core.o.core_terminate_o):
+ comb += dbg.terminate_i.eq(1)
# pass the prefix mode from Fetch to Issue, so the latter can loop
# on VL==0
# Issue is where the VL for-loop # lives. the ready/valid
# signalling is used to communicate between the four.
- # set up Fetch FSM
- fetch = FetchFSM(self.allow_overlap, self.svp64_en,
- self.imem, core_rst, pdecode2, cur_state,
- dbg, core,
- dbg.state.svstate, # combinatorially same
- nia, is_svp64_mode)
- m.submodules.fetch = fetch
- # connect up in/out data to existing Signals
- comb += fetch.p.i_data.pc.eq(dbg.state.pc) # combinatorially same
- comb += fetch.p.i_data.msr.eq(dbg.state.msr) # combinatorially same
- # and the ready/valid signalling
- comb += fetch_pc_o_ready.eq(fetch.p.o_ready)
- comb += fetch.p.i_valid.eq(fetch_pc_i_valid)
- comb += fetch_insn_o_valid.eq(fetch.n.o_valid)
- comb += fetch.n.i_ready.eq(fetch_insn_i_ready)
+ self.fetch_fsm(m, dbg, core, core_rst, nia, is_svp64_mode,
+ fetch_pc_o_ready, fetch_pc_i_valid,
+ fetch_insn_o_valid, fetch_insn_i_ready)
self.issue_fsm(m, core, nia,
dbg, core_rst, is_svp64_mode,
exec_insn_i_valid, exec_insn_o_ready,
exec_pc_o_valid, exec_pc_i_ready)
+ # whatever was done above, over-ride it if core reset is held.
+ # set NIA to pc_at_reset
+ with m.If(core_rst):
+ sync += nia.eq(self.core.pc_at_reset)
+
return m
class TestIssuer(Elaboratable):
def __init__(self, pspec):
self.ti = TestIssuerInternal(pspec)
- # XXX TODO: make this a command-line selectable option from pspec
- #from soc.simple.inorder import TestIssuerInternalInOrder
- #self.ti = TestIssuerInternalInOrder(pspec)
self.pll = DummyPLL(instance=True)
+ self.dbg_rst_i = Signal(reset_less=True)
+
# PLL direct clock or not
self.pll_en = hasattr(pspec, "use_pll") and pspec.use_pll
if self.pll_en:
# internal clock is set to selector clock-out. has the side-effect of
# running TestIssuer at this speed (see DomainRenamer("intclk") above)
# debug clock runs at coresync internal clock
- cd_coresync = ClockDomain("coresync")
- #m.domains += cd_coresync
if self.ti.dbg_domain != 'sync':
cd_dbgsync = ClockDomain("dbgsync")
- #m.domains += cd_dbgsync
- intclk = ClockSignal("coresync")
+ intclk = ClockSignal(self.ti.core_domain)
dbgclk = ClockSignal(self.ti.dbg_domain)
# XXX BYPASS PLL XXX
# XXX BYPASS PLL XXX
# XXX BYPASS PLL XXX
if self.pll_en:
comb += intclk.eq(self.ref_clk)
+ assert self.ti.core_domain != 'sync', \
+ "cannot set core_domain to sync and use pll at the same time"
else:
- comb += intclk.eq(ClockSignal())
+ if self.ti.core_domain != 'sync':
+ comb += intclk.eq(ClockSignal())
if self.ti.dbg_domain != 'sync':
dbgclk = ClockSignal(self.ti.dbg_domain)
comb += dbgclk.eq(intclk)
+ comb += self.ti.dbg_rst_i.eq(self.dbg_rst_i)
return m
}
pspec = TestMemPspec(ldst_ifacetype='bare_wb',
imem_ifacetype='bare_wb',
- addr_wid=48,
+ addr_wid=64,
mask_wid=8,
reg_wid=64,
units=units)
projects / soc.git / blobdiff