and (as long as there was no exception) the data comes out (at any
time from the PortInterface), and is captured by the LDCompSTUnit.
+TODO: dcbz, yes, that's going to be complicated, has to be done
+ with great care, to detect the case when dcbz is set
+ and *not* expect to read any data, just the address.
+ so, wait for RA but not RB.
+
Both LD and ST may request that the address be computed from summing
operand1 (src[0]) with operand2 (src[1]) *or* by summing operand1 with
the immediate (from the opcode).
* A third FSM activates to cover ST. it activates if op_is_st is true
+ * TODO document DCBZ (not complete yet)
+
* The "overall" (fourth) FSM coordinates the progression and completion
of the three other FSMs, firing "WR_RESET" which switches off "busy"
from nmigen.compat.sim import run_simulation
from nmigen.cli import verilog, rtlil
-from nmigen import Module, Signal, Mux, Cat, Elaboratable, Array, Repl
+from nmigen import Module, Signal, Mux, Cat, Elaboratable, Array, Repl, C
from nmigen.hdl.rec import Record, Layout
from nmutil.latch import SRLatch, latchregister
from soc.fu.ldst.ldst_input_record import CompLDSTOpSubset
from openpower.decoder.power_decoder2 import Data
from openpower.consts import MSR
+from soc.config.test.test_loadstore import TestMemPspec
# for debugging dcbz
from nmutil.util import Display
Data (outputs)
--------------
- * :data_o: Dest out (LD) - managed by wr[0] go/req
+ * :o_data: Dest out (LD) - managed by wr[0] go/req
* :addr_o: Address out (LD or ST) - managed by wr[1] go/req
* :exc_o: Address/Data Exception occurred. LD/ST must terminate
TODO: use one module for the byte-reverse as it's quite expensive in gates
"""
- def __init__(self, pi=None, rwid=64, awid=48, opsubset=CompLDSTOpSubset,
+ def __init__(self, pi=None, rwid=64, awid=64, opsubset=CompLDSTOpSubset,
debugtest=False, name=None):
super().__init__(rwid)
self.awid = awid
self.pi = pi
self.cu = cu = LDSTCompUnitRecord(rwid, opsubset, name=name)
- self.debugtest = debugtest
+ self.debugtest = debugtest # enable debug output for unit testing
# POWER-compliant LD/ST has index and update: *fixed* number of ports
self.n_src = n_src = 3 # RA, RB, RT/RS
- self.n_dst = n_dst = 2 # RA, RT/RS
+ self.n_dst = n_dst = 3 # RA, RT/RS, CR0
# set up array of src and dest signals
for i in range(n_src):
self.oper_i = cu.oper_i
self.src_i = cu._src_i
- self.data_o = Data(self.data_wid, name="o") # Dest1 out: RT
+ self.o_data = Data(self.data_wid, name="o") # Dest1 out: RT
self.addr_o = Data(self.data_wid, name="ea") # Addr out: Update => RA
+ self.cr_o = Data(4, name="cr0") # CR0 (for stdcx etc)
self.exc_o = cu.exc_o
self.done_o = cu.done_o
self.busy_o = cu.busy_o
#####################
# latches for the FSM.
- m.submodules.opc_l = opc_l = SRLatch(sync=False, name="opc")
+ m.submodules.opc_l = opc_l = SRLatch(sync=True, name="opc")
m.submodules.src_l = src_l = SRLatch(False, self.n_src, name="src")
m.submodules.alu_l = alu_l = SRLatch(sync=False, name="alu")
m.submodules.adr_l = adr_l = SRLatch(sync=False, name="adr")
m.submodules.sto_l = sto_l = SRLatch(sync=False, name="sto")
m.submodules.wri_l = wri_l = SRLatch(sync=False, name="wri")
m.submodules.upd_l = upd_l = SRLatch(sync=False, name="upd")
+ m.submodules.cr0_l = cr0_l = SRLatch(sync=False, name="cr0")
m.submodules.rst_l = rst_l = SRLatch(sync=False, name="rst")
m.submodules.lsd_l = lsd_l = SRLatch(sync=False, name="lsd") # done
# opcode decode
op_is_ld = Signal(reset_less=True)
op_is_st = Signal(reset_less=True)
+ op_is_dcbz = Signal(reset_less=True)
+ op_is_st_or_dcbz = Signal(reset_less=True)
+ op_is_atomic = Signal(reset_less=True)
# ALU/LD data output control
alu_valid = Signal(reset_less=True) # ALU operands are valid
rda_any = Signal(reset_less=True) # any read for address ops
rd_done = Signal(reset_less=True) # all *necessary* operands read
wr_reset = Signal(reset_less=True) # final reset condition
+ canceln = Signal(reset_less=True) # cancel (active low)
+ store_done = Signal(reset_less=True) # store has been actioned
# LD and ALU out
alu_o = Signal(self.data_wid, reset_less=True)
reset_o = Signal(reset_less=True) # reset opcode
reset_w = Signal(reset_less=True) # reset write
reset_u = Signal(reset_less=True) # reset update
+ reset_c = Signal(reset_less=True) # reset cr0
reset_a = Signal(reset_less=True) # reset adr latch
reset_i = Signal(reset_less=True) # issue|die (use a lot)
reset_r = Signal(self.n_src, reset_less=True) # reset src
reset_s = Signal(reset_less=True) # reset store
- comb += reset_i.eq(issue_i | self.go_die_i) # various
- comb += reset_o.eq(self.done_o | self.go_die_i) # opcode reset
- comb += reset_w.eq(self.wr.go_i[0] | self.go_die_i) # write reg 1
- comb += reset_u.eq(self.wr.go_i[1] | self.go_die_i) # update (reg 2)
- comb += reset_s.eq(self.go_st_i | self.go_die_i) # store reset
- comb += reset_r.eq(self.rd.go_i | Repl(self.go_die_i, self.n_src))
- comb += reset_a.eq(self.go_ad_i | self.go_die_i)
+ # end execution when a terminating condition is detected:
+ # - go_die_i: a speculative operation was cancelled
+ # - exc_o.happened: an exception has occurred
+ terminate = Signal()
+ comb += terminate.eq(self.go_die_i | self.exc_o.happened)
+
+ comb += reset_i.eq(issue_i | terminate) # various
+ comb += reset_o.eq(self.done_o | terminate) # opcode reset
+ comb += reset_w.eq(self.wr.go_i[0] | terminate) # write reg 1
+ comb += reset_u.eq(self.wr.go_i[1] | terminate) # update (reg 2)
+ comb += reset_c.eq(self.wr.go_i[2] | terminate) # cr0 (reg 3)
+ comb += reset_s.eq(self.go_st_i | terminate) # store reset
+ comb += reset_r.eq(self.rd.go_i | Repl(terminate, self.n_src))
+ comb += reset_a.eq(self.go_ad_i | terminate)
p_st_go = Signal(reset_less=True)
sync += p_st_go.eq(self.st.go_i)
# decode bits of operand (latched)
oper_r = CompLDSTOpSubset(name="oper_r") # Dest register
- comb += op_is_st.eq(oper_r.insn_type == MicrOp.OP_STORE) # ST
- comb += op_is_ld.eq(oper_r.insn_type == MicrOp.OP_LOAD) # LD
- comb += Display("compldst_multi: op_is_dcbz = %i",
- (oper_r.insn_type == MicrOp.OP_DCBZ))
+ comb += op_is_st.eq(oper_r.insn_type == MicrOp.OP_STORE) # ST
+ comb += op_is_ld.eq(oper_r.insn_type == MicrOp.OP_LOAD) # LD
+ comb += op_is_dcbz.eq(oper_r.insn_type == MicrOp.OP_DCBZ) # DCBZ
+ comb += op_is_atomic.eq(oper_r.reserve) # atomic LR/SC
+ comb += op_is_st_or_dcbz.eq(op_is_st | op_is_dcbz)
+ # dcbz is special case of store
+ #uncomment if needed
+ #comb += Display("compldst_multi: op_is_dcbz = %i",
+ # (oper_r.insn_type == MicrOp.OP_DCBZ))
op_is_update = oper_r.ldst_mode == LDSTMode.update # UPDATE
op_is_cix = oper_r.ldst_mode == LDSTMode.cix # cache-inhibit
comb += self.load_mem_o.eq(op_is_ld & self.go_ad_i)
# - alu_l : looks after add of src1/2/imm (EA)
# - adr_l : waits for add (EA)
# - upd_l : waits for adr and Regfile (port 2)
+ # - cr0_l : waits for Rc=1 and CR0 Regfile (port 3)
# - src_l[2] : ST
# - lod_l : waits for adr (EA) and for LD Data
# - wri_l : waits for LD Data and Regfile (port 1)
# opcode latch - inverted so that busy resets to 0
# note this MUST be sync so as to avoid a combinatorial loop
# between busy_o and issue_i on the reset latch (rst_l)
- sync += opc_l.s.eq(issue_i) # XXX NOTE: INVERTED FROM book!
- sync += opc_l.r.eq(reset_o) # XXX NOTE: INVERTED FROM book!
+ comb += opc_l.s.eq(issue_i) # XXX NOTE: INVERTED FROM book!
+ comb += opc_l.r.eq(reset_o) # XXX NOTE: INVERTED FROM book!
# src operand latch
- sync += src_l.s.eq(Repl(issue_i, self.n_src))
+ sync += src_l.s.eq(Repl(issue_i, self.n_src) & ~self.rdmaskn)
sync += src_l.r.eq(reset_r)
+ #### sync += Display("reset_r = %i",reset_r)
# alu latch. use sync-delay between alu_ok and valid to generate pulse
comb += alu_l.s.eq(reset_i)
#self.done_o | (self.pi.busy_o & op_is_update),
self.n_dst))
+ # CR0 operand latch (CR0 written to reg 3 if Rc=1)
+ op_is_rc1 = self.oper_i.rc.rc & self.oper_i.rc.ok
+ comb += cr0_l.s.eq(issue_i & op_is_rc1)
+ sync += cr0_l.r.eq(reset_c)
+
# update-mode operand latch (EA written to reg 2)
sync += upd_l.s.eq(reset_i)
sync += upd_l.r.eq(reset_u)
# store latch
- comb += sto_l.s.eq(addr_ok & op_is_st)
+ comb += sto_l.s.eq(addr_ok & op_is_st_or_dcbz)
sync += sto_l.r.eq(reset_s | p_st_go)
# ld/st done. needed to stop LD/ST from activating repeatedly
# create a latch/register for the operand
with m.If(self.issue_i):
sync += oper_r.eq(self.oper_i)
- with m.If(self.done_o):
+ with m.If(self.done_o | terminate):
sync += oper_r.eq(0)
- # and for LD
+ # and for LD and store-done
ldd_r = Signal(self.data_wid, reset_less=True) # Dest register
latchregister(m, ldd_o, ldd_r, ld_ok, name="ldo_r")
+ # store actioned, communicate through CR0 (for atomic LR/SC)
+ latchregister(m, self.pi.store_done.data, store_done,
+ self.pi.store_done.ok,
+ name="std_r")
+
# and for each input from the incoming src operands
srl = []
for i in range(self.n_src):
# now do the ALU addr add: one cycle, and say "ready" (next cycle, too)
comb += alu_o.eq(src1_or_z + src2_or_imm) # actual EA
- m.d.sync += alu_ok.eq(alu_valid) # keep ack in sync with EA
+ m.d.sync += alu_ok.eq(alu_valid & canceln) # keep ack in sync with EA
############################
# Control Signal calculation
# 1st operand read-request only when zero not active
# 2nd operand only needed when immediate is not active
- slg = Cat(op_is_z, op_is_imm)
+ slg = Cat(op_is_z, op_is_imm) #is this correct ?
bro = Repl(self.busy_o, self.n_src)
- comb += self.rd.rel_o.eq(src_l.q & bro & ~slg & ~self.rdmaskn)
+ comb += self.rd.rel_o.eq(src_l.q & bro & ~slg)
# note when the address-related read "go" signals are active
comb += rda_any.eq(self.rd.go_i[0] | self.rd.go_i[1])
# alu input valid when 1st and 2nd ops done (or imm not active)
- comb += alu_valid.eq(busy_o & ~(self.rd.rel_o[0] | self.rd.rel_o[1]))
+ comb += alu_valid.eq(busy_o & ~(self.rd.rel_o[0] | self.rd.rel_o[1]) &
+ canceln)
# 3rd operand only needed when operation is a store
comb += self.rd.rel_o[2].eq(src_l.q[2] & busy_o & op_is_st)
comb += self.adr_rel_o.eq(alu_valid & adr_l.q & busy_o)
# the write/store (etc) all must be cancelled if an exception occurs
- cancel = Signal(reset_less=True)
- comb += cancel.eq(self.exc_o.happened | self.shadown_i)
+ # note: cancel is active low, like shadown_i,
+ # while exc_o.happpened is active high
+ comb += canceln.eq(~self.exc_o.happened & self.shadown_i)
# store release when st ready *and* all operands read (and no shadow)
- comb += self.st.rel_o.eq(sto_l.q & busy_o & rd_done & op_is_st &
- cancel)
+ # dcbz is special case of store -- TODO verify shadows
+ comb += self.st.rel_o.eq(sto_l.q & busy_o & rd_done & op_is_st_or_dcbz &
+ canceln)
# request write of LD result. waits until shadow is dropped.
comb += self.wr.rel_o[0].eq(rd_done & wri_l.q & busy_o & lod_l.qn &
- op_is_ld & cancel)
+ op_is_ld & canceln)
# request write of EA result only in update mode
comb += self.wr.rel_o[1].eq(upd_l.q & busy_o & op_is_update &
- alu_valid & cancel)
+ alu_valid & canceln)
+
+ # request write of CR0 result only in reserve and Rc=1
+ comb += self.wr.rel_o[2].eq(cr0_l.q & busy_o & op_is_atomic &
+ alu_valid & canceln)
# provide "done" signal: select req_rel for non-LD/ST, adr_rel for LD/ST
comb += wr_any.eq(self.st.go_i | p_st_go |
- self.wr.go_i[0] | self.wr.go_i[1])
- comb += wr_reset.eq(rst_l.q & busy_o & cancel &
- ~(self.st.rel_o | self.wr.rel_o[0] |
- self.wr.rel_o[1]) &
- (lod_l.qn | op_is_st)
+ self.wr.go_i.bool())
+ comb += wr_reset.eq(rst_l.q & busy_o & canceln &
+ ~(self.st.rel_o | self.wr.rel_o.bool()) &
+ (lod_l.qn | op_is_st_or_dcbz)
)
comb += self.done_o.eq(wr_reset & (~self.pi.busy_o | op_is_ld))
# Data/Address outputs
# put the LD-output register directly onto the output bus on a go_write
- comb += self.data_o.data.eq(self.dest[0])
+ comb += self.o_data.data.eq(self.dest[0])
+ comb += self.o_data.ok.eq(self.wr.rel_o[0])
with m.If(self.wr.go_i[0]):
comb += self.dest[0].eq(ldd_r)
# "update" mode, put address out on 2nd go-write
comb += self.addr_o.data.eq(self.dest[1])
+ comb += self.addr_o.ok.eq(self.wr.rel_o[1])
with m.If(op_is_update & self.wr.go_i[1]):
comb += self.dest[1].eq(addr_r)
+ # fun-fun-fun, calculate CR0 when Rc=1 requested.
+ cr0 = self.dest[2]
+ comb += self.cr_o.data.eq(cr0)
+ comb += self.cr_o.ok.eq(self.wr.rel_o[2])
+ with m.If(cr0_l.q):
+ comb += cr0.eq(Cat(C(0, 1), store_done, C(0, 2)))
+
# need to look like MultiCompUnit: put wrmask out.
# XXX may need to make this enable only when write active
- comb += self.wrmask.eq(bro & Cat(op_is_ld, op_is_update))
+ comb += self.wrmask.eq(bro & Cat(op_is_ld, op_is_update, cr0_l.q))
###########################
# PortInterface connections
# connect to LD/ST PortInterface.
comb += pi.is_ld_i.eq(op_is_ld & busy_o) # decoded-LD
- comb += pi.is_st_i.eq(op_is_st & busy_o) # decoded-ST
+ comb += pi.is_nc.eq(op_is_cix & busy_o) # cache-inhibited
+ comb += pi.is_st_i.eq(op_is_st_or_dcbz & busy_o) # decoded-ST
+ comb += pi.is_dcbz_i.eq(op_is_dcbz & busy_o) # decoded-DCBZ
+ comb += pi.reserve.eq(oper_r.reserve & busy_o) # atomic LR/SC
comb += pi.data_len.eq(oper_r.data_len) # data_len
# address: use sync to avoid long latency
sync += pi.addr.data.eq(addr_r) # EA from adder
+ with m.If(op_is_dcbz):
+ sync += Display("LDSTCompUnit.DCBZ: EA from adder %x", addr_r)
+
sync += pi.addr.ok.eq(alu_ok & lsd_l.q) # "do address stuff" (once)
comb += self.exc_o.eq(pi.exc_o) # exception occurred
comb += addr_ok.eq(self.pi.addr_ok_o) # no exc, address fine
- # connect MSR.PR for priv/virt operation
- comb += pi.msr_pr.eq(oper_r.msr[MSR.PR])
+ # connect MSR.PR etc. for priv/virt operation
+ comb += pi.priv_mode.eq(~oper_r.msr[MSR.PR])
+ comb += pi.virt_mode.eq(oper_r.msr[MSR.DR])
+ comb += pi.mode_32bit.eq(~oper_r.msr[MSR.SF])
+ with m.If(self.issue_i): # display this only once
+ sync += Display("LDSTCompUnit: oper_r.msr %x pr=%x dr=%x sf=%x",
+ oper_r.msr,
+ oper_r.msr[MSR.PR],
+ oper_r.msr[MSR.DR],
+ oper_r.msr[MSR.SF])
# byte-reverse on LD
revnorev = Signal(64, reset_less=True)
comb += pi.st.data.eq(stdata_r)
with m.Else():
comb += pi.st.data.eq(op3)
+
# store - data goes in based on go_st
comb += pi.st.ok.eq(self.st.go_i) # go store signals st data valid
to LDSTOutputData o and o1 respectively.
"""
if i == 0:
- return self.data_o # LDSTOutputData.regspec o
+ return self.o_data # LDSTOutputData.regspec o
if i == 1:
return self.addr_o # LDSTOutputData.regspec o1
+ if i == 2:
+ return self.cr_o # LDSTOutputData.regspec cr_a
# return self.dest[i]
def get_fu_out(self, i):
yield self.adr_rel_o
yield self.sto_rel_o
yield self.wr.rel_o
- yield from self.data_o.ports()
+ yield from self.o_data.ports()
yield from self.addr_o.ports()
+ yield from self.cr_o.ports()
yield self.load_mem_o
yield self.stwd_mem_o
yield dut.src1_i.eq(src1)
yield dut.src2_i.eq(src2)
yield dut.src3_i.eq(src3)
- yield dut.oper_i.imm_data.imm.eq(imm)
+ yield dut.oper_i.imm_data.data.eq(imm)
yield dut.oper_i.imm_data.ok.eq(imm_ok)
- yield dut.oper_i.update.eq(update)
+ #guess: this one was removed -- yield dut.oper_i.update.eq(update)
yield dut.issue_i.eq(1)
yield
yield dut.issue_i.eq(0)
if rel == active_rel:
break
yield
- yield dut.rd.go.eq(active_rel)
+ yield dut.rd.go_i.eq(active_rel)
yield
- yield dut.rd.go.eq(0)
+ yield dut.rd.go_i.eq(0)
yield from wait_for(dut.adr_rel_o, False, test1st=True)
# yield from wait_for(dut.adr_rel_o)
yield dut.src1_i.eq(src1)
yield dut.src2_i.eq(src2)
yield dut.oper_i.zero_a.eq(zero_a)
- yield dut.oper_i.imm_data.imm.eq(imm)
+ yield dut.oper_i.imm_data.data.eq(imm)
yield dut.oper_i.imm_data.ok.eq(imm_ok)
yield dut.issue_i.eq(1)
yield
# wait for the operands (RA, RB, or both)
if rd:
- yield dut.rd.go.eq(rd)
+ yield dut.rd.go_i.eq(rd)
yield from wait_for(dut.rd.rel_o)
- yield dut.rd.go.eq(0)
+ yield dut.rd.go_i.eq(0)
yield from wait_for(dut.adr_rel_o, False, test1st=True)
# yield dut.ad.go.eq(1)
if update:
yield from wait_for(dut.wr.rel_o[1])
- yield dut.wr.go.eq(0b10)
+ yield dut.wr.go_i.eq(0b10)
yield
addr = yield dut.addr_o
print("addr", addr)
- yield dut.wr.go.eq(0)
+ yield dut.wr.go_i.eq(0)
else:
addr = None
yield from wait_for(dut.wr.rel_o[0], test1st=True)
- yield dut.wr.go.eq(1)
+ yield dut.wr.go_i.eq(1)
yield
- data = yield dut.data_o
- print(data)
- yield dut.wr.go.eq(0)
+ data = yield dut.o_data.o
+ data_ok = yield dut.o_data.o_ok
+ yield dut.wr.go_i.eq(0)
yield from wait_for(dut.busy_o)
yield
# wait_for(dut.stwd_mem_o)
- return data, addr
+ return data, data_ok, addr
def ldst_sim(dut):
class TestLDSTCompUnit(LDSTCompUnit):
- def __init__(self, rwid):
+ def __init__(self, rwid, pspec):
from soc.experiment.l0_cache import TstL0CacheBuffer
- self.l0 = l0 = TstL0CacheBuffer()
- pi = l0.l0.dports[0].pi
+ self.l0 = l0 = TstL0CacheBuffer(pspec)
+ pi = l0.l0.dports[0]
LDSTCompUnit.__init__(self, pi, rwid, 4)
def elaborate(self, platform):
m = LDSTCompUnit.elaborate(self, platform)
m.submodules.l0 = self.l0
- m.d.comb += self.ad.go.eq(self.ad.rel) # link addr-go direct to rel
+ # link addr-go direct to rel
+ m.d.comb += self.ad.go_i.eq(self.ad.rel_o)
return m
def test_scoreboard():
- dut = TestLDSTCompUnit(16)
+ units = {}
+ pspec = TestMemPspec(ldst_ifacetype='bare_wb',
+ imem_ifacetype='bare_wb',
+ addr_wid=64,
+ mask_wid=8,
+ reg_wid=64,
+ units=units)
+
+ dut = TestLDSTCompUnit(16,pspec)
vl = rtlil.convert(dut, ports=dut.ports())
with open("test_ldst_comp.il", "w") as f:
f.write(vl)
class TestLDSTCompUnitRegSpec(LDSTCompUnit):
- def __init__(self):
+ def __init__(self, pspec):
from soc.experiment.l0_cache import TstL0CacheBuffer
from soc.fu.ldst.pipe_data import LDSTPipeSpec
regspec = LDSTPipeSpec.regspec
- self.l0 = l0 = TstL0CacheBuffer()
- pi = l0.l0.dports[0].pi
+ self.l0 = l0 = TstL0CacheBuffer(pspec)
+ pi = l0.l0.dports[0]
LDSTCompUnit.__init__(self, pi, regspec, 4)
def elaborate(self, platform):
m = LDSTCompUnit.elaborate(self, platform)
m.submodules.l0 = self.l0
- m.d.comb += self.ad.go.eq(self.ad.rel) # link addr-go direct to rel
+ # link addr-go direct to rel
+ m.d.comb += self.ad.go_i.eq(self.ad.rel_o)
return m
def test_scoreboard_regspec():
- dut = TestLDSTCompUnitRegSpec()
+ units = {}
+ pspec = TestMemPspec(ldst_ifacetype='bare_wb',
+ imem_ifacetype='bare_wb',
+ addr_wid=64,
+ mask_wid=8,
+ reg_wid=64,
+ units=units)
+
+ dut = TestLDSTCompUnitRegSpec(pspec)
vl = rtlil.convert(dut, ports=dut.ports())
with open("test_ldst_comp.il", "w") as f:
f.write(vl)
projects / soc.git / blobdiff