self.regreduce_en = (hasattr(pspec, "regreduce") and
(pspec.regreduce == True))
+ # and if overlap requested
+ self.allow_overlap = (hasattr(pspec, "allow_overlap") and
+ (pspec.allow_overlap == True))
+
# 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
# pulse to synchronize the simulator at instruction end
self.insn_done = Signal()
+ # indicate any instruction still outstanding, in execution
+ self.any_busy = Signal()
+
if self.svp64_en:
# store copies of predicate masks
self.srcmask = Signal(64)
self.dstmask = Signal(64)
- def fetch_fsm(self, m, core, pc, svstate, nia, is_svp64_mode,
+ def fetch_fsm(self, m, dbg, core, pc, svstate, nia, is_svp64_mode,
fetch_pc_o_ready, fetch_pc_i_valid,
fetch_insn_o_valid, fetch_insn_i_ready):
"""fetch FSM
# waiting (zzz)
with m.State("IDLE"):
- comb += fetch_pc_o_ready.eq(1)
+ with m.If(~dbg.stopping_o):
+ comb += fetch_pc_o_ready.eq(1)
with m.If(fetch_pc_i_valid):
# instruction allowed to go: start by reading the PC
# capture the PC and also drop it into Insn Memory
# dummy pause to find out why simulation is not keeping up
with m.State("INSN_READ"):
- # one cycle later, msr/sv read arrives. valid only once.
- with m.If(~msr_read):
- sync += msr_read.eq(1) # yeah don't read it again
- sync += cur_state.msr.eq(self.state_r_msr.o_data)
- with m.If(self.imem.f_busy_o): # zzz...
- # busy: stay in wait-read
- comb += self.imem.a_i_valid.eq(1)
- comb += self.imem.f_i_valid.eq(1)
+ if self.allow_overlap:
+ stopping = dbg.stopping_o
+ else:
+ stopping = Const(0)
+ with m.If(stopping):
+ # stopping: jump back to idle
+ m.next = "IDLE"
with m.Else():
- # not busy: instruction fetched
- insn = get_insn(self.imem.f_instr_o, cur_state.pc)
- if self.svp64_en:
- svp64 = self.svp64
- # decode the SVP64 prefix, if any
- comb += svp64.raw_opcode_in.eq(insn)
- comb += svp64.bigendian.eq(self.core_bigendian_i)
- # pass the decoded prefix (if any) to PowerDecoder2
- sync += pdecode2.sv_rm.eq(svp64.svp64_rm)
- sync += pdecode2.is_svp64_mode.eq(is_svp64_mode)
- # remember whether this is a prefixed instruction, so
- # the FSM can readily loop when VL==0
- sync += is_svp64_mode.eq(svp64.is_svp64_mode)
- # calculate the address of the following instruction
- insn_size = Mux(svp64.is_svp64_mode, 8, 4)
- sync += nia.eq(cur_state.pc + insn_size)
- with m.If(~svp64.is_svp64_mode):
- # with no prefix, store the instruction
- # and hand it directly to the next FSM
+ # one cycle later, msr/sv read arrives. valid only once.
+ with m.If(~msr_read):
+ sync += msr_read.eq(1) # yeah don't read it again
+ sync += cur_state.msr.eq(self.state_r_msr.o_data)
+ with m.If(self.imem.f_busy_o): # zzz...
+ # busy: stay in wait-read
+ comb += self.imem.a_i_valid.eq(1)
+ 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)
+ if self.svp64_en:
+ svp64 = self.svp64
+ # decode the SVP64 prefix, if any
+ comb += svp64.raw_opcode_in.eq(insn)
+ comb += svp64.bigendian.eq(self.core_bigendian_i)
+ # pass the decoded prefix (if any) to PowerDecoder2
+ sync += pdecode2.sv_rm.eq(svp64.svp64_rm)
+ sync += pdecode2.is_svp64_mode.eq(is_svp64_mode)
+ # remember whether this is a prefixed instruction,
+ # so the FSM can readily loop when VL==0
+ sync += is_svp64_mode.eq(svp64.is_svp64_mode)
+ # calculate the address of the following instruction
+ insn_size = Mux(svp64.is_svp64_mode, 8, 4)
+ sync += nia.eq(cur_state.pc + insn_size)
+ with m.If(~svp64.is_svp64_mode):
+ # with no prefix, store the instruction
+ # and hand it directly to the next FSM
+ sync += dec_opcode_i.eq(insn)
+ m.next = "INSN_READY"
+ with m.Else():
+ # fetch the rest of the instruction from memory
+ comb += self.imem.a_pc_i.eq(cur_state.pc + 4)
+ comb += self.imem.a_i_valid.eq(1)
+ comb += self.imem.f_i_valid.eq(1)
+ m.next = "INSN_READ2"
+ else:
+ # not SVP64 - 32-bit only
+ sync += nia.eq(cur_state.pc + 4)
sync += dec_opcode_i.eq(insn)
m.next = "INSN_READY"
- with m.Else():
- # fetch the rest of the instruction from memory
- comb += self.imem.a_pc_i.eq(cur_state.pc + 4)
- comb += self.imem.a_i_valid.eq(1)
- comb += self.imem.f_i_valid.eq(1)
- m.next = "INSN_READ2"
- else:
- # not SVP64 - 32-bit only
- sync += nia.eq(cur_state.pc + 4)
- sync += dec_opcode_i.eq(insn)
- m.next = "INSN_READY"
with m.State("INSN_READ2"):
with m.If(self.imem.f_busy_o): # zzz...
# note if an exception happened. in a pipelined or OoO design
# this needs to be accompanied by "shadowing" (or stalling)
- el = []
- for exc in core.fus.excs.values():
- el.append(exc.happened)
- exc_happened = Signal()
- if len(el) > 0: # at least one exception
- comb += exc_happened.eq(Cat(*el).bool())
+ exc_happened = self.core.o.exc_happened
with m.FSM(name="issue_fsm"):
# pass predicate mask bits through to satellite decoders
# TODO: for SIMD this will be *multiple* bits
- sync += core.sv_pred_sm.eq(self.srcmask[0])
- sync += core.sv_pred_dm.eq(self.dstmask[0])
+ sync += core.i.sv_pred_sm.eq(self.srcmask[0])
+ sync += core.i.sv_pred_dm.eq(self.dstmask[0])
# after src/dst step have been updated, we are ready
# to decode the instruction
with m.State("DECODE_SV"):
# decode the instruction
- sync += core.e.eq(pdecode2.e)
- sync += core.state.eq(cur_state)
- sync += core.raw_insn_i.eq(dec_opcode_i)
- sync += core.bigendian_i.eq(self.core_bigendian_i)
+ sync += core.i.e.eq(pdecode2.e)
+ sync += core.i.state.eq(cur_state)
+ sync += core.i.raw_insn_i.eq(dec_opcode_i)
+ sync += core.i.bigendian_i.eq(self.core_bigendian_i)
if self.svp64_en:
- sync += core.sv_rm.eq(pdecode2.sv_rm)
+ sync += core.i.sv_rm.eq(pdecode2.sv_rm)
# set RA_OR_ZERO detection in satellite decoders
- sync += core.sv_a_nz.eq(pdecode2.sv_a_nz)
+ sync += core.i.sv_a_nz.eq(pdecode2.sv_a_nz)
# and svp64 detection
- sync += core.is_svp64_mode.eq(is_svp64_mode)
+ sync += core.i.is_svp64_mode.eq(is_svp64_mode)
# and svp64 bit-rev'd ldst mode
ldst_dec = pdecode2.use_svp64_ldst_dec
- sync += core.use_svp64_ldst_dec.eq(ldst_dec)
+ sync += core.i.use_svp64_ldst_dec.eq(ldst_dec)
# after decoding, reset any previous exception condition,
# allowing it to be set again during the next execution
sync += pdecode2.ldst_exc.eq(0)
pdecode2 = self.pdecode2
# temporaries
- core_busy_o = core.busy_o # core is busy
- core_ivalid_i = core.ivalid_i # instruction is valid
- core_issue_i = core.issue_i # instruction is issued
- insn_type = core.e.do.insn_type # instruction MicroOp type
+ core_busy_o = core.n.o_data.busy_o # core is busy
+ core_ivalid_i = core.p.i_valid # instruction is valid
with m.FSM(name="exec_fsm"):
with m.State("INSN_START"):
comb += exec_insn_o_ready.eq(1)
with m.If(exec_insn_i_valid):
- comb += core_ivalid_i.eq(1) # instruction is valid
- comb += core_issue_i.eq(1) # and issued
+ comb += core_ivalid_i.eq(1) # instruction is valid/issued
sync += sv_changed.eq(0)
sync += pc_changed.eq(0)
- m.next = "INSN_ACTIVE" # move to "wait completion"
+ with m.If(core.p.o_ready): # only move if accepted
+ m.next = "INSN_ACTIVE" # move to "wait completion"
# instruction started: must wait till it finishes
with m.State("INSN_ACTIVE"):
- with m.If(insn_type != MicrOp.OP_NOP):
- comb += core_ivalid_i.eq(1) # instruction is valid
# note changes to PC and SVSTATE
with m.If(self.state_nia.wen & (1<<StateRegs.SVSTATE)):
sync += sv_changed.eq(1)
comb += dbg_rst.eq(ResetSignal())
# busy/halted signals from core
- comb += self.busy_o.eq(core.busy_o)
+ core_busy_o = ~core.p.o_ready | core.n.o_data.busy_o # core is busy
+ comb += self.busy_o.eq(core_busy_o)
comb += pdecode2.dec.bigendian.eq(self.core_bigendian_i)
# temporary hack: says "go" immediately for both address gen and ST
pc_changed = Signal() # note write to PC
sv_changed = Signal() # note write to SVSTATE
+ # indicate to outside world if any FU is still executing
+ comb += self.any_busy.eq(core.n.o_data.any_busy_o) # any FU executing
+
# read state either from incoming override or from regfile
# TODO: really should be doing MSR in the same way
pc = state_get(m, core_rst, self.pc_i,
# connect up debug signals
# TODO comb += core.icache_rst_i.eq(dbg.icache_rst_o)
- comb += dbg.terminate_i.eq(core.core_terminate_o)
+ comb += dbg.terminate_i.eq(core.o.core_terminate_o)
comb += dbg.state.pc.eq(pc)
comb += dbg.state.svstate.eq(svstate)
comb += dbg.state.msr.eq(cur_state.msr)
# Issue is where the VL for-loop # lives. the ready/valid
# signalling is used to communicate between the four.
- self.fetch_fsm(m, core, pc, svstate, nia, is_svp64_mode,
+ self.fetch_fsm(m, dbg, core, pc, svstate, nia, is_svp64_mode,
fetch_pc_o_ready, fetch_pc_i_valid,
fetch_insn_o_valid, fetch_insn_i_ready)